15 Commits

Include renames

Author	SHA1	Message	Date
t0saki	b7d0415c24	feat(plugins): skill catalog and openviking-skills for the memory plugins (#5161) ... * refactor(skills): install skills through one shared helper POST /api/v1/skills kept its whole install loop (source resolution, per-skill install, source metadata, list_only) inline in the route. Move it into openviking/server/skill_ingest.py:install_skills so the MCP add_skill tool and signed skill uploads can reuse the exact same code path. The REST route's behavior is unchanged. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(mcp): add an add_skill tool MCP clients had no way to create a skill: write refuses the skills/ subtree (_USER_MANAGED_SUBTREES) and add_resource validates its target as a resource. Agents that should keep skills in OpenViking could read them but never add one. add_skill takes either the full SKILL.md text (data) or a path. A Git or GitHub tree URL installs through the same source resolution as REST, with skills=[...] to pick from a multi-skill repository and list_only to preview it. A local SKILL.md, directory, or zip gets the add_resource treatment: the tool mints a one-time upload token, now tagged kind="skill" with the target root, selection and list_only, and the signed temp_upload installs the file as skills instead of ingesting it as a resource. target_uri="viking://agent/skills" shares the skill with the account. All three paths (REST, MCP inline/Git, signed upload) go through skill_ingest.install_skills. The tool count in the server log, app comment, docs, and the Codex plugin's REAL_MCP_TOOLS moves to 16. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(mcp): search shared skills in find(context_type="skill") Without a target_uri, find resolved the generic default targets, which stop at the caller's user root, so a skill search never reached the account-shared viking://agent/skills. REST /skills/find and the context search already cover both roots. When context_type resolves to skill only and no target_uri is given, the MCP tool now targets default_target_directories(ctx, context_type=SKILL): the user's own skills plus viking://agent/skills. REST find semantics are unchanged. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(mcp): print directory abstracts in tree(include_abstract=true) The tree tool skipped to the next entry right after printing a directory, and only printed abstracts for files, but the storage layer only fills abstracts for directories (files always come back empty). The flag therefore never printed anything. Print the abstract after either kind of entry and ask for up to 1024 characters, enough for a full skill description, so tree(uri="viking://~/skills", level_limit=1, include_abstract=true) lists every skill with its description. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(skills): honor node_limit in GET /api/v1/skills list_skills declared node_limit but always listed each skill root with a hardcoded 1000. Pass it through per root; 0 keeps the default so the CLI's accepted range (-n 0) still lists everything. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(mcp): point skill hits in find/search at their SKILL.md A skill is indexed through its directory's .abstract.md, so find and list-mode search printed hits like viking://agent/skills/x/.abstract.md. Following the "use the read tool to expand a URI" advice returned only the frontmatter, and read_content inlined the same stub. Skill hits now show <dir>/SKILL.md, and read_content reads that file. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(mcp): validate add_skill targets and sources before minting an upload Review findings on the add_skill tool: - target_uri passed the content-kind check for any path under a skills root (viking://~/skills/pdf) and, for ROOT, for another user's root, but the installer only accepts the caller's own skills root or viking://agent/skills. On the local-path branch the tool minted a one-time upload token anyway, and the upload failed with 400 after the token was spent. The target is now resolved with the installer's own rule first; shared subpaths map to viking://agent/skills, the rest fail at once, and the error names both allowed roots. - Non-Git remote sources such as tos:// were treated as remote, then refused as "direct host filesystem paths". add_skill now decides Git with the same prefixes resolve_skill_source uses (shared as GIT_SKILL_SOURCE_PREFIXES) and reports other schemes as unsupported. - With list_only, the upload instructions still said the skill would be installed and that no further call was needed; they now say the upload only lists the source's skills. - The zip example packaged hidden files, so .git and .env files went into the stored skill. It now excludes VCS data, .env files, node_modules and .DS_Store, starting from a fresh archive. - tree(include_abstract=true) printed the "abstract is not ready" placeholder for directories that never get an abstract, such as a skill's scripts/. Those placeholders are skipped. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * docs(mcp): say that write only refuses the user's own skills subtree The capability reference claimed MCP write refuses every skill URI. It refuses the user's own skills/ subtree, but under viking://agent/skills it writes a plain file that skips skill installation. State that, and point shared skills at add_skill as well. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * style(skills): format skill_processor.py Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(plugins): inject an <available-skills> catalog at session start Agents on every harness learned about memories at session start but had no idea which skills OpenViking held, so a stored skill was only found if a later recall happened to surface it. buildProfileBlock() now takes the caller's resolved config as a fourth argument and, when skillCatalog is on (default), adds <available-skills> after <available-memories>: one GET /api/v1/skills lists the user's own skills first, then account-shared ones, dropping a shared skill the user shadows by name. Descriptions are cut to about 40 tokens and envelope tags in them are escaped, since the shared root is written by anyone on the account. The block has its own budget (skillCatalogTokenBudget, default 1200) and degrades from descriptions to names to a one-line count; with no skills, or a server without the endpoint, it is omitted. The shared formatListing now gives entries back so its "+N more" tail fits: a greedily filled listing never left room for it, so a cut listing ended silently. This also applies to <available-memories>. All six callers (claude-code, codex, opencode, dsh, pi, and the thin hook runtime for cursor/trae/zcode) pass their config through. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(plugins): recall account-shared skills and flag skill entries The server context face already mixes both skill roots into per-prompt recall, but the plugins' last-resort ranked find only searched viking://~/memories and viking://~/skills, so on servers without the context face a shared skill in viking://agent/skills could never be recalled. Add it as a third source, and name each skill hit by its directory rather than the .abstract.md it was indexed through, matching the context face and the session-start catalog. When an injected recall block carries a skill (a type="skills" entry, or a [skill] line from the fallback), its header gains one line telling the agent to read the skill's SKILL.md before following it. Turns without a skill are unchanged. The openclaw plugin's vendored recall-core copy is regenerated with it. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(plugins): point skill writes at add_skill in the URI guard A local Write or Edit aimed at viking://.../skills/... was denied with a hint to use MCP write or edit instead, but the server refuses both under the skills subtree, so the hint led straight into a second error. Hints may now depend on the URI: for a skill URI (viking://~/skills, viking://user/<id>/skills, viking://agent/skills) the default table and dsh's bridged table name add_skill with an add_skill(data="<the full SKILL.md text>") example. Other URIs keep their hints. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(plugins): bundle an openviking-skills skill Nothing told an agent how to work with skills stored in OpenViking: how to load one from the catalog, run its helper files, create one through add_skill, install from Git or a local folder, share it with the account, or move the user's existing local skills over. examples/skills/openviking-skills covers all of that, including a user-triggered, one-time migration of ~/.claude/skills, ~/.agents/skills and ~/.cursor/skills that keeps environment-bound skills local (shipped by a plugin or marketplace, symlinked in by a CLI installer such as lark-cli, or needing a local binary) and uploads only what the user approves skill by skill. It passes strict server validation. sync.mjs ships it wherever add_skill is a real tool and a bundled skill loads: the codex, claude-code, cursor and dsh plugins. openviking-memory now points to it for skill work. A new sync test keeps synced skills flat, since copySkill copies a flat file list and a subdirectory would crash it; the marketplace tests pin the packaged copies, and dsh's provider test expects both bundled skills. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * chore(plugins): bump versions for the skill integration claude-code 0.6.0, codex 0.10.0, agent-hook (cursor/trae/zcode) 0.4.0 and dsh 0.5.0 gain the skill catalog and, except trae/zcode, the bundled openviking-skills skill; opencode 0.3.3 and pi 0.3.3 gain the catalog. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(codex): recall shared skills and flag skill entries in Codex too Codex builds its recall block itself instead of through recall-core's wrappers, so the previous commit's changes never reached it: its raw search still skipped viking://agent/skills, its digest carried no skill hint, and its post-processing kept only level-2 leaves, which silently dropped every skill hit (skills are found through their directory's level-0 abstract), including the viking://~/skills search it already ran. recall-core now exports skillEntryHint() and skillHitUri(). The hint is added when a block carries a type="skills" entry or cites any skill URI, which also covers digests that only keep URIs. Codex searches the shared skill root as a third bucket, labels skill hits "skills" under their directory URI, lets them through post-processing, and adds the same hint line to its envelope. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(cursor): install openviking-skills next to openviking-memory sync.mjs puts openviking-skills into hosts/cursor/skills, but the installer copied only openviking-memory into ~/.cursor/skills, so Cursor never saw the new skill. Install, uninstall, the post-install check and the doctor's file list now cover both skills. The install test also moves to the agent-hook plugin's new 0.4.0 version string. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(dsh): keep PLUGIN_VERSION in step with package.json The version bump moved package.json to 0.5.0 but left the PLUGIN_VERSION constant at 0.4.3, which bundle.test.mjs and npm run check:version compare against the manifest. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * docs(skills): match ov-skills commands to the real ov CLI flags The ov-skills skill documented flags the CLI never had (--json, and a --limit that is only a hidden alias), a raw-content "ov skills add -" form that sends a literal "-", and ov resources subcommands that do not exist. Every command line now follows the clap definitions: -o json for JSON, -n/--node-limit, -p/--uri on read commands and -p/--parent-auto-create on add, -s/--skill as a comma list, show --format, and validate's --strict-only body-length warning. ov add-skill is documented as the same command as ov skills add. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * docs(plugins): document the skill catalog, openviking-skills, and skill-aware recall The integration pages (Claude Code, Codex, Cursor, TRAE, opencode, pi, dsh), the capability reference, the plugin development guide, and the plugin READMEs now describe the <available-skills> session-start block, its skillCatalog / skillCatalogTokenBudget knobs, the bundled openviking-skills skill where it ships, recall reaching viking://agent/skills with the skill-entry hint, and the URI guard sending skill writes to add_skill. en and zh pages carry the same facts. Stale statements fixed on the lines touched: thin hook hosts use the same 10000-token profile budget as the rest, the claude-code and codex plugins ship four skills, dsh mounts the server MCP surface, and the opencode install guide lists openviking_add_skill once. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(plugins): let the user's own skills use the catalog budget the shared group leaves Review findings on <available-skills>: - Each group got at most its even share of the listing budget, with unused tokens passing only forward, so the user's own skills (always first) never got more than half. Twenty own skills and one shared skill fell back to names only while most of the 1200 tokens went unused. A group now takes its even share or everything the later groups leave when listed in full, whichever is larger. - When a group's share could not hold its header plus the "+N more" tail, formatListing gave back every entry and printed a bare header, which reads as an empty directory. It now prints the one-line "N entries, budget too tight" stub instead (memory listings too). - A budget too small for even the one-line count now injects nothing rather than overrunning it. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(plugins): rank skill hits like memory leaves in the recall fallbacks Naming a skill hit by its directory instead of its .abstract.md cost it the 0.12 leaf boost, since the boost keyed on a ".md" URI, so a skill that main would recall lost to ten slightly weaker memory leaves. In Codex, skills were also never picked while enough memory leaves passed the threshold (leaves are picked first), and a hit the server labeled with another category lost its "skills" label. Skill hits now count as leaves in ranking and picking in both recall-core and Codex, and Codex always labels them "skills". Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(plugins): send shared-skill edits back to the shared root in the URI guard The guard's add_skill example carried no target_uri, and add_skill without one installs into the caller's own root. Fixing a shared skill that way left the team copy unchanged and created a private copy that the catalog then shows instead. For viking://agent/skills URIs the example now passes target_uri="viking://agent/skills", and a helper file (anything below a skill's SKILL.md) points to a folder upload through add_skill(path=...) rather than SKILL.md text. addSkillExample() builds the example for both the default table and dsh's. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * docs(skills): tighten openviking-skills where review found unsafe steps - The catalog is a snapshot that drops descriptions or entries with many skills, so a name missing from it does not prove the name is free. Check <root>/<name>/SKILL.md, and confirm with the user before replacing an existing skill, since add_skill replaces silently. - Updating a shared skill must pass target_uri="viking://agent/skills" after the user confirms; otherwise add_skill creates a private copy that shadows it. - Every file in an uploaded folder is stored with the skill: zip without .git, .env files, node_modules and .DS_Store, and delete the archive afterwards. The migration now inspects the whole folder, hidden files included, for secrets. - Migration flags frontmatter keys OpenViking drops (for example disable-model-invocation or context), and fixes a missing name or description in a temporary copy, never in the user's file. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * feat(plugins): keep the session-start block under the host's inline limit Claude Code saves hook context over 10,000 characters to a file and shows the model a 2 KB preview; Codex and trae-cli spill past about 10,000 bytes; ZCode drops stdout over 32 KB. The session-start block (profile at a 10,000-token budget, memory index, skill catalog, and the archive on resume) routinely ran 25-40 KB, so on these hosts the model saw only the start of the profile, and the catalog appended at the end never reached it. sessionStartMaxBytes caps the whole block in UTF-8 bytes: 9500 for claude-code and codex, 20000 for zcode, no cap elsewhere. Under the cap buildProfileBlock shrinks its token budgets (about 4 bytes per estimated token); if the block still does not fit it drops the memory index, then the catalog. On resume/compact the archive takes up to half, truncated on a line with a pointer to viking://~/sessions/<id>/history/. On resume, claude-code and codex skip the profile block when it matches the one this session already received, since the restored history holds it; a changed block is injected again. Claude-Session: https://claude.ai/code/session_01ECVkFufAU2LKe4g83cxt28 * fix(mcp): return one hit per skill package in find #5045 made a skill index as a whole package, so an item-level find now returns one hit per file inside it. Route skill-only find through SearchService.find_skills, which keeps the best hit per package, and resolve every skill hit to its package's SKILL.md instead of only rewriting the two index sidecars. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(mcp): point skill changes at add_skill in the tool descriptions The server keeps accepting write/edit under viking://agent/skills, and forget still removes a skill directory, so the constraint lives in the tool descriptions: add_skill is the one entry point for creating and updating a skill, and removal goes through ov skills remove or Studio. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * fix(mcp): describe a skill hit by its own abstract A package hit can be any file inside the skill, whose abstract describes that file and not the skill, so find would list a skill under a helper script's summary. Read the package's abstract for those hits, the way GET /skills/find already does. Keep a filter-only skill query on the generic find, which find_skills does not serve. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(mcp): document package-level skill retrieval in find Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * feat(agent-plugins): ship the openviking-skills skill Agent Plugins has no hooks, so no session-start catalog: without this skill the model never learns that the account's skills exist. The skill's own text now reaches for find(context_type="skill") first and treats <available-skills> as something only some harnesses inject, so one copy reads correctly in both kinds of harness. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * fix(plugins): name the skill package a recall hit came from Since a skill is indexed as a whole package, a hit can be any file inside it, not just the two index sidecars the old rewrite stripped. Derive the package root the way the server's skill_root_uri does, drop the internal update backups, and keep one entry per package at its best score. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * fix(plugins): let the skill catalog use both roots' full listings The server already caps each skill root at node_limit, so a second cap over the merged list only bites once the private root alone fills it — and then it drops the shared root whole while reporting nothing dropped. The token budget is what should decide, and it already does. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(plugins): say node_limit caps each skill root, not the merged list Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * fix(mcp): do not paste an unready abstract over a skill hit's own summary fs.abstract returns a placeholder string rather than raising when a package has no usable .abstract.md, so the substitution replaced a useful file summary with a diagnostic line. Reject the same placeholders tree already rejects, and bound the per-package reads the way read_content is bounded. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(mcp): correct how search reports skill hits, and refresh the tool table find and search both render one line per skill package, so the earlier wording — that search returns several hits per package — contradicted the code. Say what actually differs: search still spends a limit slot per matching file and keeps that file's summary. Also point forget and add_resource at add_skill where an agent would look for them, name the REST delete alongside the CLI, and bring the capability table's line citations back in step with mcp_endpoint.py. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(agent-plugins): list add_skill among the tools the package exposes Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * refactor(mcp): drop guards and prose no caller can reach install_skills only ever returns a dict, add_skill always fills root_uri, and a source with no SKILL.md raises before it gets here, so the isinstance, empty-list and missing-uri branches were unreachable. fs.abstract only returns the directory placeholder. One skill package resolves to one rendered item, so the pending map holds one each. In the docstrings, drop what Args already says and the one removal path an MCP caller cannot take. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(plugins): drop a comment about a branch formatListing cannot take A listing left with only its header returns the stub above, so it never reaches the silent close the comment described. Also name sessionStartMaxBytes in buildProfileBlock's options type, where the .d.mts already has it. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * revert(plugins): drop the skill changes in the recall fallback Reverts the recall half of the skill integration: 992215bfb, 553416200, 5e838a0a0 and 62a456c08. The seven files they touched go back to their state at aa77061c1, the merge base with main. Those four commits taught the local recall assembly about skills: a third source for viking://agent/skills, skillHitUri naming a hit's package, dedupeSkillHits keeping one entry per package, rankItem scoring a skill like a memory leaf, and a header line telling the model to read SKILL.md. Four of the five only ever ran in the raw-find fallback. recallForPeer calls buildServerAssembledBlock first and returns as soon as it answers, so searchAllSources, rankItem and the fallback block builder are reached only on a deployment whose server has no context face. The fifth, the header line in wrapContext, did run on the main path, but the server already reports each entry's type and the URI it wants read, so the line restates what the block carries. Skills still reach the model on the path that runs: the context face searches both skill roots, returns them as entries with type="skills", and the session-start <available-skills> catalog lists every skill with its description. Both stay. The documentation that described the fallback behaviour goes with it. The statements that survive are the ones the context face makes true on its own, such as recall covering the skills shared with the account. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * chore(plugins): bump opencode and pi past main's releases Both were 0.3.3 on this branch and main has since shipped 0.3.3 of its own, so the version a host installs by no longer moves for the skill catalog they now carry through the shared library. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW * docs(mcp): say what list-mode search actually reports for a skill hit The search row claimed a skill package's summary is the matching file's. It is not: _format_search_result rewrites every skill hit onto the package's SKILL.md, keeps the best-scored one per package, and _describe_skills_by_package replaces the summary with the package's own abstract. What is true is that limit applies during retrieval, before that merge, so a package matching several files still spends several slots and fewer than limit results come back. Claude-Session: https://claude.ai/code/session_01CrZadR75kCueyZyoiGUFBW	2026-09-22 15:09:44 +08:00
t0sakiandTRAE CLI	aa3ee2f431	refactor(plugins): resolve every harness's knobs from one declaration, and generate the packaged copies at pack time (#4773) ... * fix(codex): stop dropping turns on an outage, and honour bypass patterns Two capabilities every other JS harness has were wired in the shared library but never reached codex. Offline queue. `catchUpTurns` called `sendSessionMessages` without `enqueueOnRetryable`, so a 5xx or a connection failure returned a count and the turns were gone: codex hooks are short-lived subprocesses with no retry of their own, and the `capturedTurnCount` cursor only compensates while the process survives long enough for another Stop. The queue module was vendored into `scripts/shared/` all along with no caller, and nothing ever replayed it. The send now enqueues, and SessionStart drains the queue after its health check — the one codex hook that runs against a known-healthy server. Bypass. `isBypassed` had no codex caller, so `bypassSessionPatterns` did nothing there: a directory a user had excluded still had its turns captured and still got memories injected, and the `bypass.session_patterns` key a repository commits in `.openviking/config.json` was silently inert for every codex user in that repository. All five hooks now consult the shared matcher, and the loader reads the same keys and the same `OPENVIKING_BYPASS_SESSION[_PATTERNS]` env vars as claude-code. SessionStart is deliberately only half-suppressed under bypass: it skips peer registration and injection, but still replays the queue and still sweeps. Both finish work recorded by sessions that were not bypassed, and this hook is the only place codex runs either, so bailing out would strand that data for as long as the user kept working in a bypassed repository. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * fix(pi): use the shared bypass matcher, and keep the pre-git peer reachable `config.ts` kept only `.peerId` out of `resolveEffectivePeerId()` and dropped `legacyPeerId`, so under `recallPeerScope: "actor"` — where the effective peer is the only one asked — every memory written before the git-derived peer replaced the path-derived one became unreachable. The shared recall builder has read `options.legacyPeerId` for the dual read all along; pi just never passed it. Bypass was a hand-written `matchBypass()` under this extension's own `bypassPatterns` key: it understood a leading or trailing `*` and nothing else, so `**/scratch/**` did not work here while it did everywhere else. It now calls the shared `isBypassed`, reads `bypassSessionPatterns` like every other harness, and honours `OPENVIKING_BYPASS_SESSION[_PATTERNS]`. `bypassPatterns` still reads. That does change one behaviour for existing setups: a bare path used to match its subdirectories as a prefix, and a glob does not. The README says to write `"/tmp/scratch**"` where `"/tmp/scratch"` used to be enough. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * chore(plugins): remove the deprecated standalone TRAE CLI integration `examples/trae-cli-memory-hooks` was added on 2026-08-17 (#4026) for TraeCode CLI 1.0 and deprecated the next day by #4079, which routed `--harness trae-cli` to the codex plugin alias instead. Since then `install_trae_cli` has not been in the installer's dispatch block — `normalize_trae_cli_harness` rewrites the harness to `codex` before it could run — so the directory, its 588 lines and its CI test have been dead weight that the marketplace archive still shipped to every user. The uninstall path stays: `remove_legacy_trae_cli_integration` and `agent_remove_trae_cli_configs` delete what an old install left on disk and never read this directory, which is what the retention test in `release-marketplace.test.mjs` claimed to protect. `install-agent-hooks.test.mjs` covers that path and still passes. Also drops the write half that only `install_trae_cli` called. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * fix(plugins): make a changed plugin prove it shipped Claude Code resolves an installed plugin by the `version` string in its manifest, not by the commit the marketplace ref points at. That string has read 0.4.5 since #4389 on 2026-08-27, with seven commits to the plugin and the shared library behind it — so nobody running the marketplace build has received any of them. Codex's manifest has been frozen at 0.8.1 for ten. Every vendored-copy diff those commits paid for went to a payload no user received. Both versions move, and a new PR check fails when files under a marketplace-distributed plugin change without its version string moving with them. The generated `shared/` copies live inside each plugin directory, so a shared-library change reaches the check through them. Also removes `claude-code-memory-plugin/package-lock.json`: 1174 lines locking 95 packages for a private manifest that declares no dependencies, still pinned at 0.4.4. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * docs(agent-integrations): record what codex, pi and trae-cli now do The capability matrices said codex has no on-disk queue and counted four hooks where hooks.json declares five; both are now wrong. pi's bypass is no longer a prefix match under its own key. And the removed TRAE CLI 1.0 integration is described by what remains of it: the installer's uninstall path. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * refactor(plugins): derive each plugin's vendored set from what it imports sync.mjs kept hand-written capability groups, and the groups got spread into each other: the file's own header states the rule it could not enforce ("what a plugin ships must equal what it imports"). Both directions had failed in practice — a module a plugin imported but no list named became ERR_MODULE_NOT_FOUND on the first hook of a fresh install, and modules nobody imported sat vendored in plugin trees, landing in every review diff forever. The lists are gone. Each target now says only where its code lives and where its copies go; the file set is the transitive closure of the imports that code actually resolves into the vendored directory — including re-export shims like claude-code's scripts/lib/, which are the sole importer of several modules. The sync reports a copy nothing imports and a lib module no plugin reaches, and sync.test.mjs asserts what is on disk equals that closure in both directions. The run is byte-identical for everything still shipped, and it named the two copies nobody had noticed: claude-code's capture-utils (515 lines, its auto-capture hand-rolls the same filter) and codex's uri-guard (78 lines, codex has no PreToolUse hook). ZCode stops vendoring altogether. Its only install path is `assemble_agent_integration`, the same one cursor and trae use, which already placed the canonical runtime beside it — it just imported its own committed copy instead. Pointing its nine imports at `../../memory-plugin-shared/lib/`, the path that resolves both in the repository and under `$OV_HOME/agent-integrations/`, drops 4294 generated lines and leaves cursor, trae and zcode reading one runtime. The installer's assembly list grows by the three modules zcode needs, derived by the same closure code so it cannot drift either, and the end-to-end archive install test covers it. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * fix(plugins): run cursor and trae capture through the shared filter zcode got this in #4594; cursor and trae never did. Both sent whatever their transcript parser produced straight to the extractor: `/compact`, a bare "ok", a punctuation-only turn and a `[openviking-memory]` status line all became memories, and a turn past `captureMaxLength` went out uncapped. Every other harness has run `shouldCaptureText` on the way in for months. The decision now lives beside the runtime the thin harnesses compose, as `filterCaptureTurns`, so a harness gets it by calling rather than by carrying its own copy — and the installer already assembles `capture-utils.mjs` for these three since they share one runtime. Both hooks keep hashing the raw turn rather than the filtered text, so raising captureMaxLength never resends a turn the server already holds truncated, and a dropped turn is recorded as handled so a re-read of the same transcript does not re-evaluate it every run. The TRAE hook test used `last_assistant_message: "done"` — an acknowledgement the filter drops, and correctly so. Its fixture now carries a turn worth remembering. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * fix(plugins): bump every plugin the shared change reached, and pair the manifests The version check added earlier in this branch caught what a human review would not have: `filterCaptureTurns` landed in the shared library, its generated copy changed inside opencode and dsh, and neither manifest moved — so both hosts would have kept installing the build without it. cursor, trae and pi move for the same reason. It also caught a mismatch this branch introduced: cursor carries the version in both `.cursor-plugin/plugin.json` and `openviking.integration.json`, the host installs by the first and the installer decides "nothing changed" by the second, and bumping only one makes a plugin report upgraded while behaving as it did. zcode had been sitting at 0.1.1 against 0.1.2 on main for the same reason. The check now compares the pair, so neither can drift again. Claude-Session: https://claude.ai/code/session_01GJouP1ghK4gwXw7sDjuZX6 * refactor(plugins): resolve every harness's knob from one declaration A setting only stayed uniform across the harnesses when a shared module read it on the hot path. `recallLimit`, `scoreThreshold` and `captureMaxLength` agree everywhere because `recall-core` and `capture-utils` read them; the switches, which nothing shared read, drifted into four spellings — a boolean `autoRecall`, an `autoRecall` object in opencode, `syncTurns` in dsh and pi, and no recall switch at all in the thin hook runtime. Three more inventories had to be kept in step by hand: the doctor's known-knob set, the workspace file's dotted-key map, and the implicit list inside `loadAgentHookConfig`. `lib/config-schema.mjs` is now the one declaration, and the rest are projections of it. `resolveSettings()` resolves any harness through the layers the shared README has documented since the `plugin` section landed — env, the workspace file and registry, `ovcli.conf` `plugin.<harness>`, `ovcli.conf` `plugin`, ov.conf's harness section, defaults — which only Claude Code and Codex implemented. cursor, trae, trae-cn and zcode read the environment and nothing else before this, so a `plugin` entry named after them was inert and `ov config switch` moved their credentials while leaving their behaviour behind. The two per-plugin config files are gone with it. pi's `config.json` shipped with the extension holding exactly the code defaults, so nothing could tell an operator's choice from the factory setting; opencode searched four paths for `openviking-config.json`. Both harnesses read `ovcli.conf` now, as the others do. The `.d.mts` stubs move to lib/ beside the modules they describe — pi and openclaw each carried a different, both incomplete, declaration of `recall-core` — and sync ships them to the targets written in TypeScript. The switches move into the capability modules that own them: `isRecallEnabled` and `isCaptureEnabled` accept every spelling and treat any of them saying off as off, which is the one mechanism in this repository that has actually held a name still. * build(plugins): generate the packaged plugins' shared copies at pack time `sync.mjs` had zero references under `.github/`: the generator ran on whoever last remembered, and CI only checked the result byte for byte. So every change to a 6,000-line library arrived as a 20,000-line diff, and the copies were committed for plugins that never needed them committed. The split is delivery, not taste. A host that installs Claude Code, Codex or the Agent Plugins package points at a directory in this repository and can only load what git holds, so those copies stay committed — and a push to main regenerates them, which is what stops one going stale behind a merged pull request. opencode, dsh and openclaw publish as npm packages and pi is tarred by the installer, so those four build their copies on the way out: a `prepack` script for the packages, the marketplace staging script for the archive the installer reads, and `sync.mjs` in a source checkout for a direct `install.sh` run. `npm pack` was verified to regenerate a deleted directory and ship the same file set. Two things had been holding those directories in place. Their release triggers matched only `examples/<plugin>/**`, so a shared fix reached npm through the vendored copy changing — the triggers name the library now. The version gate found a changed plugin the same way, so it treats a change under `memory-plugin-shared/lib` as a change to every plugin, and gains cursor and trae, which are distributed the same way and were never gated. pi's own `recall-ledger` moves to `lib/`, where the extension's other own modules live, so the generated directory holds nothing but generated files. * test(plugins): pin the layer order for the harnesses that had no config test cursor, trae, trae-cn and zcode compose one shared loader and none of them owns a config test, so nothing would have caught the `plugin` section going inert again. This asserts the whole stack through that loader — the shared block, a per-harness override under either spelling, the workspace file over ovcli.conf, and the environment over all of it — and pins that a default never reports as configured, since several fields reach the server only when the user asked for them. * docs(plugins): point opencode and pi at ovcli.conf, and say when copies are made Both plugins documented a configuration file that no longer exists, down to the four paths opencode searched for it and the nested blocks neither loader reads any more. Their knobs live in `ovcli.conf`'s `plugin` section now, so the examples are ovcli.conf examples, the flattened names are the ones the schema declares, and each document says the resolution order and links to the one file that declares every knob. Two claims that had gone the other way: opencode and pi read the workspace `.openviking/config.json` now, so a `peer.id` written there does apply, and the shared README says which plugins keep their generated copies in git and which build them at pack time, because a fresh checkout has to run the generator once before their tests will resolve an import. * fix(opencode): name the file the 401 hint should send the user to * fix(plugins): let the canonical knob name win over its own alias A file that carries both spellings — what a rename leaves behind — resolved to whichever the layer loop reached last, which was the older name. * fix(plugins): write each generated copy through a rename The marketplace staging script runs the generator now, so it can rewrite a vendored module while a test is byte-comparing it. A partial file read that way fails as a drifted copy, which is a confusing way to say nothing drifted. * fix(dsh,pi): honour the recall switch these two never read Both gained a capture switch when `syncTurns` became an alias, but their recall paths still retrieved on every prompt: nothing in either called the switch, so `autoRecall: false` — settable from ovcli.conf, a workspace file or the environment — was accepted everywhere and obeyed nowhere. * docs(agent-integrations): say which harnesses read which configuration layer The capability reference had cursor, trae, trae-cn and zcode down as environment-variable-only, opencode reading a four-level `openviking-config.json` search and pi its own `config.json`, and the workspace files and `plugin` section scoped to Claude Code and Codex. All nine resolve through the same loader now, so the tables, the quick-scope list and the profile cards say so, and the module table gains the schema the layers project from. The distribution paragraph was counting modules against a `HOOK_SHARED_FILES` list that no longer exists and calling zcode the only harness that vendors the hook runtime, which stopped being true when its copies were dropped. The counts are the real ones, and the paragraph now says when each target's copies are generated, since that is what decides whether git holds them. * build(plugins): keep node_modules out of the marketplace archive Three of the staged plugin trees carry installed development dependencies, so the copy shipped 134MB nobody unpacks and spent most of the release test's 54 seconds on them. Copying through tar leaves them behind while still carrying the generated shared copies, which a tracked-file listing would miss. * build(plugins): read the installer's shared-module list from the sync The installer carried a second, hand-written copy of the closure sync.mjs already computes, and only a regex over the bash text held the two equal. The sync writes that closure to lib/MANIFEST now and the installer copies what it names, so the set has one source and the test compares two lists instead of scraping a script. The installer cannot just run the sync: a marketplace archive is flat, and the generator resolves an empty list there. * build(plugins): derive what the marketplace archive must contain The stage script hand-listed 56 required paths beside three generators that already knew them, and it had drifted from all three: eleven of the twenty-two shared modules the installer assembles, four of the thirty-eight copies the sync generates, and neither marketplace manifest. The list is now computed from each plugin's own manifests, the imports those entrypoints reach, and the sync's target list, leaving the script with only the directories it ships. * fix(plugins): resolve credentials from the calling harness's ov.conf section The shared resolver read `ov.conf`'s `codex` block for every caller, so the api key, account, user and peer of opencode, pi, dsh, cursor, trae, trae-cn and zcode all came out of a block named after another harness, while the block named after them did nothing. A user who wrote `opencode.apiKey` saw no effect and a stale `codex` block quietly authenticated everything else. `resolveOpenVikingCredentials` now takes the harness and reads the section named after it, under either spelling. The layer keeps its place — after ovcli.conf, ahead of `server.root_api_key` — and the ovcli-pinned mode still skips ov.conf entirely, so codex, the caller that keeps the default, is unchanged. For users this means the `<harness>` block finally configures that harness, and anyone who was relying on the `codex` block to supply credentials to a different harness has to copy them into their own block. dsh merges its section in explicitly because the cordis patch occupies its legacy layer, and codex now honours an api key set through `ovcli.conf`'s `plugin` section, which the credential chain has never seen. cursor, trae, trae-cn and zcode honour it too: the hook runtime spread the resolved credentials over the settings, so an empty key from the chain used to erase the one `plugin.<harness>.apiKey` had supplied. The doctor warns about all of these blocks now, since the server refuses to start on any of them. * fix(plugins): let the credential chain supply the peer on claude-code and dsh Claude Code had no peer logic at all: it took only the user agent from the shared credential module, so `actor_peer_id` in ovcli.conf and `peerId` in ov.conf's `claude_code` block were read by every other harness and ignored here. `ov config switch` moved the peer and Claude Code kept writing under the one derived from the workspace. It now takes the peer from the same chain, and a `plugin` entry or a workspace file still names the more specific answer for that directory, which is the rule the thin harnesses already follow. dsh had the opposite fault: it assigned the credential peer over whatever the layers resolved, and that chain is the empty string whenever no actor peer is named, so `plugin.dsh.peerId` was resolved and then thrown away. The credential peer is a fallback now instead of an overwrite. * fix(plugins): resolve the peer through one chain on every harness Six loaders each wrote their own peer order, and two of them disagreed with the rest: codex let a `peerId` written for the harness win, while opencode and pi let `ovcli.conf`'s account-wide `actor_peer_id` win. One ovcli.conf therefore produced two different peers depending on which host read it, and `ov config switch` moved the peer for some harnesses and not others. `resolvePluginPeerId()` in the shared library is now that order, once: a peer the host named, then `OPENVIKING_PEER_ID` (suppressed when the credentials are pinned to ovcli.conf, where the environment is meant not to apply), then the workspace file, the registry and `ovcli.conf`'s `plugin` section, then `ovcli.conf`'s `actor_peer_id`, and last `ov.conf`'s harness block, which keeps the place it has always had. Nothing named means nothing explicit, which is what makes `peer.source` derive one. For opencode and pi users this is a precedence change: a `peerId` under `plugin` or `plugin.<harness>` in ovcli.conf, or a `peer.id` in a workspace file, now outranks that file's `actor_peer_id` instead of being outranked by it. Anyone who wrote both and wanted the actor peer has to drop the plugin one. It is one for claude-code too: `ov.conf`'s `claude_code.peerId` was the only peer that harness read, and it now sits under `actor_peer_id` like every other harness's block. Otherwise nothing moves: `ov.conf`'s harness block is ranked here rather than left to the credential chain, which drops it whenever the credentials are pinned to ovcli.conf, so that block names the peer in the same place whether they are pinned or not. * fix(plugins): shape the cursor and trae MCP proxies like every other one Both proxies hand-built the config object the shared `buildMcpProxyConfig` produces everywhere else, and had done so since before that helper existed, so they never picked up the fixes it carries. The visible one is the actor peer: they put whatever peer the config layers resolved straight onto `X-OpenViking-Actor-Peer`, while the default `recallPeerScope` of `all` means broad recall and every other proxy deliberately sends no such header. Cursor and TRAE therefore recalled at a narrower scope than the rest on identical configuration. Two smaller ones come with it: the watch list now includes the default credential paths, so a proxy started before the first `ov login` picks up the file it creates instead of never reloading, and the request timeout is clamped rather than taken raw. The guard is the file list in `mcp-proxy-config.test.mjs`, which named five proxies by hand and so never covered the two that were wrong. It is a directory scan now, pinned to a count so a rename cannot quietly empty it, and it asserts that every proxy in the tree reaches the shared builder. * fix(plugins): put one set of headers on the wire from every harness Seven request builders each wrote their own header block and three of them disagreed. Codex sent the api key twice, as `Authorization: Bearer` and again as `X-API-Key`, which the open-source server prefers when both arrive — so a gateway rewriting one of them changed which credential authenticated. Claude Code read `cfg.accountId` where the other six read `cfg.account`. And only Codex asked whether the server was in trusted mode before naming the operator: everywhere else `X-OpenViking-Account` / `X-OpenViking-User` went out whenever they resolved, including to `api_key` servers that read both out of the key, ignore the headers, and leave the identity visible to every proxy on the path. Both doctors have been warning users about that as if it were already fixed. On the wire, from this commit: - `X-API-Key` stops. Codex's `ov-session.mjs`, `auto-recall.mjs` and `session-start-commit.mjs` were the only senders; `Authorization: Bearer` is unchanged and remains the only credential header. A gateway that needs `X-API-Key` has to add it itself. openclaw keeps sending it and is untouched. - `X-OpenViking-Account` and `X-OpenViking-User` are sent only under `sendIdentityHeaders`, which is `authMode === "trusted"`. Codex already behaved this way; claude-code, opencode, dsh, pi and the cursor / trae / trae-cn / zcode hook runtime now do too, and so do the two senders outside the hook stacks: every harness's stdio MCP proxy, which takes the switch through `buildMcpProxyConfig` the way it already takes the actor peer, and Claude Code's status-line server probe. Auth mode resolves through `resolveAuthMode()` in the shared credentials module: `plugin.<harness>` `authMode`, then `plugin.authMode`, then `ov.conf` `<harness>.authMode` on the three harnesses whose loader reads that section as a settings layer (claude-code, codex and dsh), then `ov.conf` `server.auth_mode`, then trusted whenever an account or a user resolved at all — the chain Codex already had, now shared. An operator whose server is in `api_key` mode and who has an account in ovcli.conf stops sending it; nothing else changes. - `X-OpenViking-Actor-Peer`, `User-Agent` and `Content-Type` are unchanged. Claude Code's config now also exposes the resolved identity as `account` / `user` beside the existing `accountId` / `userId`, and its doctor and its status-line probe use the headers the plugin would really send instead of always sending both. The Agent Plugins package resolves an auth mode of its own, so its proxy still names the operator to a trusted server. The free half of the same theme: `session-start-commit.mjs` carried a private `requestJSON` / `commitOvSession` pair that duplicated `ov-session.mjs`'s down to the envelope. It imports them now, which is one header block fewer to keep in step. `wire-headers.test.mjs` drives six of the seven stacks through a stubbed fetch and asserts one header map for all of them, in trusted mode and in api_key mode; `auto-recall.test.mjs` covers the seventh, which only exists as a subprocess, against a real server socket; and `mcp-proxy-config.test.mjs` puts the proxy's two identity headers on the same switch. * test(openclaw): run only the live test copies under tests/ut Six files sat directly under tests/ as copies of suites that had since moved on, and vitest's default include meant two of them still ran on every `npm test` against an older shape of the code. What each one is: - `tests/context-engine-assemble.test.ts` is an earlier cut of `tests/ut/context-engine-assemble.test.ts` — same imports one directory shallower, 360 lines against the live copy's 838. Five of its six cases are in the live copy verbatim; the sixth, "records senderId from runtimeContext in assemble diagnostics", is covered by the `senderIdFound` assertions in `tests/ut/context-engine-modules.test.ts` and `tests/ut/context-engine-afterTurn.test.ts`. - `tests/context-bloat-730.test.ts` imports `memory-ranking.js`, `config.js` and `auto-recall.js`, and every symbol it exercises already has a home under `tests/ut`: `postProcessMemories` and `pickMemoriesForInjection` in `memory-ranking.test.ts`, `buildMemoryLinesWithBudget` and `estimateTokenCount` in `build-memory-lines.test.ts`, `recallScoreThreshold` and `recallMaxInjectedChars` in `config.test.ts` and `query-config.test.ts`. - `tests/test-memory-chain.py` and `tests/test-tool-capture.py` are earlier drafts of `tests/e2e/test-memory-chain.py` and `tests/e2e/test-tool-capture.py`, which drive the same phases against a live gateway at greater length. - `tests/demo-memory-ajie.py` and `tests/demo-memory-xiaomei.py` are one manual gateway demo under two personas, differing in the user name and the port, and named by no doc, script or manifest. Nothing referenced them: the only prose pointer to an assemble test names the `tests/ut` copy, and `.clawhubignore` plus `tsconfig.build.json` already kept the whole directory out of the published package. So that a file dropped beside the suite is not silently picked up again, `vitest.config.ts` now pins `test.include` to `tests/ut/**/*.test.ts`. The CI job's `--exclude` filters still apply on top of it — 42 files listed, 41 after the architecture-boundaries exclude. The one file left in `__tests__` falls outside the pin until it moves into `tests/ut` with the rest. `architecture-boundaries.test.ts` kept `tests/context-bloat-730.test.ts` in the list of files it reads for index-facade imports, which would have thrown on the missing path, so that entry goes too. * test(openclaw): move the shouldBypassSession cases under tests/ut `vitest.config.ts` now collects only `tests/ut`, and the six cases in `__tests__/bypass-session-patterns.test.ts` were the sole behaviour coverage of `compileSessionPatterns`, `matchesSessionPattern`, `shouldBypassSession` and the `ingestReplyAssistIgnoreSessionPatterns` fallback in the config schema, so they had stopped running. They move beside the other `text-utils.ts` tests, one directory deeper, and the emptied `__tests__` directory goes with the two build files that still had to name it. * docs(openclaw): drop five reports that nothing points at any more Four are dated one-off write-ups of manual runs, and three of them name the repository and branch they were made against, neither of which is this one: - `docs/openviking-install-real-scenario-verification-report.md` (2026-06-05) walks one machine's install of plugin `2026.6.2` from `iaasng/arkclaw-openviking-plugin`, against a hosted endpoint and a specific OpenClaw build. - `docs/openviking-dynamic-query-config-test-report.md` (2026-06-04) is the acceptance run for the runtime query-config work in that same repository; `docs/openviking-runtime-query-config.md` is the reference that documents the feature and stays. - `docs/workmemory-v2-test-report.md` (2026-05-02) is a `locomo10` benchmark of Working Memory v2, beside the design note that still describes it. - `openclaw-multi-tenant-test-report.md` records a session driven through two ad-hoc remote ports and a live Feishu bot, reproducible by nobody. The fifth, `docs/oc-resource-skill-import-design.md`, is an RFC whose own opening paragraph says the implementation went the other way: no unified `ov_import(kind=...)` tool and no `/ov-import` command, which is the shape the rest of the document is about. One line pointed at any of them — the Working Memory entry in the guide's reference list, whose neighbour is the design note and stays. Nothing shipped moves: `docs/` is outside the package's `files` list, the top-level report was never in it, and the ClawHub release workflow names no document. * test(openclaw): drop the unreferenced tool-result compression bench tests/toolresult_compression_tests/ was manual measurement tooling: run-sccs-bench.mjs spawns an external openclaw binary turn by turn and reports token usage, so it asserts nothing and needs three repositories cloned to hardcoded /root paths plus a live server before it can run at all. Nothing outside the directory named it — the two invocation lines the plan cites were in its own README. The behaviour it exercised is covered by tests/ut: tool-round-trip.test.ts asserts that an externalized tool result survives conversion with its preview text, its viking://session/<id>/tool-results/<id> ref and original_chars intact, and tools.test.ts asserts the restore half through openviking_tool_result_read / _search / _list. * test(openclaw): assert the prepack script that regenerates shared at pack time The packaged plugin's prepack gained a `sync.mjs` prefix when the generated shared copies stopped being committed, but the manifest contract still pinned prepack to `npm run build` alone. The openclaw-tests job runs that file, so the suite has been red on a stale expectation rather than on a real contract break. * chore(claude-code): drop the two unwired debug scripts Neither debug-recall.mjs nor debug-capture.mjs is reachable from hooks.json, install.sh, or the sync target, so they drifted away from the hooks they were meant to mirror; debug-capture also rewrites the live session's capture cursor through an API flow the plugin no longer uses. The doctor skill, both READMEs and the capability reference lose their pointers in the same change. * docs(pi): rewrite DESIGN.md around the modules that exist The spec was written in the future tense of a plan only partly executed: it specified an `index_builder.ts` that was never built (the profile block from shared/profile-inject.mjs folded into systemPrompt took its role), gave no section at all to config.ts or takeover.ts, carried per-file line estimates that went stale on every commit, and repeated its Knowledge Index sample twice. Its Implementation Order and Testing Strategy described work the code and tests/ have both already done. It is now one section per module actually shipped, each written from that module's current code, plus the event walkthrough, the design ancestry and the two rationales the code cannot state for itself. TAKEOVER.md's model, runtime flow, compaction and failure modes move into the takeover section so the one harness capability that is unique to pi is documented beside the module that implements it; its configuration table was already in README.md. * docs(codex): fold VERIFICATION.md into a README Testing section The 320-line manual SOP was pinned to plugin v0.6.0 and every step it walked is now asserted by the node --test suite that CI runs. Only its two live legs -- extraction landing in the user namespace and the interactive Codex smoke test -- need a real server, so they stay as a short appendix next to the command that actually runs the suite. * chore(opencode): drop three unreferenced helpers from lib/utils.mjs makeMultipartRequest was a near-copy of makeRequest that no caller ever reached, and the wired URI check is lib/viking-uri-guard.mjs, not validateVikingUri; ensureRemoteUrl had no caller either. * chore(pi): drop six unreachable OVClient methods createSession, addMessageParts, addMessagePayload and deleteSession have no caller: the OV session id is derived locally and the server-side session is created implicitly by the batch messages endpoint, messages go out as one addMessage or through that batch endpoint, and nothing in the extension deletes a session. resolveScopeSpace and resolveTargetUri were a URI-space rewriter that no tool path ever reached, and they held the only readers of RESERVED_USER, RESERVED_AGENT and the resolvedSpaces cache. The sync-barrier stub kept an addMessagePayload key for the same reason; sync.ts queues and replays through fetchJSON, so the key asserted nothing. * chore(dsh): drop nine unreachable OpenVikingClient methods health, ensureSession, getSessionArchive, find, read, list, stat, forget and addResource had no caller: the runtime only ever uses fetchJSON, healthResult, ensureSessionResult, getSession, addMessage and commitSession, and the tool surface the read and write helpers look like they serve is bridged through the stdio MCP proxy, so they could never be reached. Only tests called them. The peer-override and live-recall cases move onto ensureSessionResult and healthResult, which take the same arguments, and the 409/ALREADY_EXISTS tolerance moves to a runtime test: initializeState carries its own copy of that check, so the behaviour stays pinned where it now lives. * chore(codex): drop the ov-credentials pass-through scripts/ov-credentials.mjs re-exported two names from shared/credentials.mjs and repeated that module's CLI verb-for-verb. Nothing ran it as a command, so the file only added a second path to the same resolver, one that every credential fix since the shared library landed had to remember to keep in step. Its three importers now reach the shared module directly. The test file keeps its name because .github/workflows/pr.yml lists it by path, and it asserts the shared resolver either way. * chore(install): drop the pre-stdio layout migrations The rc-wrapper stripper and the two legacy-marketplace migrations cleaned up after installs that predate the stdio MCP proxy and the unified `openviking` marketplace name, both of which landed in July 2026. Every install since then writes the current layout, so the three helpers ran on every upgrade only to find nothing, while keeping four constants and a TOML rewriter alive for a shape the installer can no longer produce. The installer-side cleanup is what goes; both doctors still detect a leftover `openviking-plugins-local` install and tell the user how to remove it by hand. The Codex README's step list loses the line that promised the migration, and the installer test that asserted a Cursor-only run left the rc blocks alone goes with the code it was pinning. * chore(claude-code): drop the dead archive-abstract loop SessionStart rendered up to five `<archive-abstract>` entries from the session context's `pre_archive_abstracts`, but the server hard-codes that field to an empty array: openviking/session/session.py returns `[]` from both context builders and says so in a comment ("保留字段返回空数组，保持 API 向下兼容"). The loop is dead twice over, since the field's entries are objects while the filter kept only strings. What resumed and compacted sessions actually receive is unchanged: the `<session-archive>` block still carries `latest_archive_overview`, which is the one field the endpoint fills. The capability reference loses the "≤5 pre_archive_abstracts" claim it made for claude-code alone. * refactor(plugins): put the four JS hook stacks on one HTTP module Claude Code, Codex, dsh and the thin-harness runtime each carried their own AbortController, header block and envelope parser for the same server, which is how the wire drifted apart in the first place. lib/ov-http.mjs owns that shape now — Bearer only, identity headers only when the config says the server is trusted with the operator's name — so the next change to it lands once instead of four times. Also fixes the usage line in credentials.mjs, which still named the wrapper that was deleted out from under it. * refactor(plugins): move the Claude Code and Codex session stacks onto the runtime Claude Code's scripts/lib/ov-session.mjs kept its own copy of the session helpers the shared hook runtime already had — a fetch builder, the pending-queue enqueue, add-message, commit, and the two session getters — so a fix to how a failed write is parked had to be made twice, and the two copies had already drifted: only Claude Code's set pendingQueued / pendingEnqueueFailed and warned about a non-retryable failure. The runtime carries that contract now. makeAgentFetchJSON takes the timeout and the actor-peer getter its callers vary (Codex knows its peer only after loading state under the session lock, and Claude Code names one per call), and resolves the workspace peer on demand so a caller that never reads it pays nothing. addAgentMessage and commitAgentSession report what became of a failed write the way the capture hooks log it, and commitAgentSession takes the retention payload Claude Code's threshold commit sends. Both harnesses' modules are wrappers over it, keeping Codex's result-or-null fetchJSON beside the envelope fetchJSONRes. That report now reaches further than Claude Code: the three thin harnesses on this runtime — cursor, trae/trae-cn and zcode — name a write no retry can fix on stderr too. * refactor(plugins): put the last header blocks on the shared builder opencode and pi still carried a whole fetch stack of their own, and the doctor, the MCP proxy, the Claude Code statusline probe and the Codex recall hook still spelled the header block out by hand. All six go through lib/ov-http.mjs now: opencode's fetchJSON and makeRequest are a shell over it that keeps its throwing contract and its two hints, pi's client keeps its typed methods over the same envelope, and the four header-only callers ask buildOvHeaders. A test asserts what that buys: X-OpenViking-Actor-Peer appears in the plugin family's non-test sources only in lib/ov-http.mjs, so the next hand-rolled header block fails before it can drift. Its scope comes from the sync targets, so a harness added there is covered without touching the test. Codex's recall hook stops overriding the per-call actor peer, which makes its legacy-peer sweep ask for the legacy peer instead of asking twice for the effective one. * refactor(claude-code): put auto-recall back on the shared recall core Claude Code's auto-recall.mjs carried a line-for-line copy of the recall core's query profile, lexical overlap boost, ranking, dedup, user-space resolution, multi-source search and budgeted block builder — 280 lines that only its own statusline numbers kept it from calling. Those numbers now come from the shared side: buildRecallBlockDetailed returns the block together with the counts the fallback builder already computed, plus a stage that says which path produced it, so a host can tell an empty recall the server had nothing to offer from one the score threshold emptied. buildRecallBlock stays the string-returning shell over it that opencode and pi call. The hook is left with the Claude Code envelope and the last-recall.json snapshot the statusline reads, and a test covers all eight reasons that file records — including a guard that fails if the ranking pipeline is ever copied back into this plugin. * chore(claude-code): drop the dead per-message capture filter auto-capture.mjs's shouldCapture has had no call site since the hook moved to batches: its length bounds, slash-command, punctuation-only and question-only rules each drop a whole multi-turn batch for what one turn in it looks like, and the comment at the batch gate has said so ever since. That left four regexes reachable only from a function nothing calls. The batch gate keeps what it always did — skip an empty batch, and in keyword mode require some user turn to carry a trigger phrase — and its comment stops naming the function that is gone. * refactor(claude-code): read the transcript through the shared capture utilities Claude Code's transcript walk was written twice — once in auto-capture, once in subagent-stop — and neither copy was the shared one every other harness uses, so a fix to block handling reached nine harnesses and skipped this one. Both now go through scripts/cc-transcript.mjs, which re-exports the shared capture utilities and adds only what is genuinely Anthropic-shaped: a tool_result nests its output in a content array and names its call by id, and that output is dropped from the turn's text while travelling verbatim in the turn's tool part. The kept turns, their order and their parts are unchanged, so the capture cursor still means what it did. The shared sanitizer now also strips <system-reminder> blocks and [Subagent Context] lines; without them this switch would have started sending Claude's own notes to itself to OpenViking as if the user had written them. * test(codex): pin the digest URI repair the shared compressor performs The recall compressor is a small model spawned through `codex exec`, and it occasionally rewrites a long viking:// URI while rephrasing a bullet, which leaves a dead link in the injected digest. Since codex went onto the shared `compressRecallContext`, every citation is snapped back onto the URIs the server actually returned before the digest is injected; this test pins that a mangled URI comes back repaired, with the short-input short-circuit disabled so the compressor is actually exercised. * refactor(plugins): open every hook entry with the shared stage The Claude Code and Codex hook entries each read stdin, re-resolved the config for the payload's directory and answered the enabled and bypass gates in their own words, and the copies had drifted: Codex's PreCompact never re-checked its own switch after the reload, and the two auto-recalls asked the two gates in opposite orders. runHookStage is that opening. An entry hands it a config loader, its gates and its envelope, and keeps only its own work, which now returns what the envelope should carry instead of writing stdout from wherever it happens to stop. The stage also carries the bypass verdict for the hook that must not stop on it — Codex's SessionStart still sweeps and replays for other sessions inside a bypassed repository — and a single-shot emit for the one that answers before its worker starts. The write-path preamble stays in the entry: a detaching hook has to spawn its worker before stdin is consumed, and only the entry knows the response its host expects. Claude Code's uri-guard keeps its own code as well; it reads no config and answers no gates. * refactor(plugins): fold the agent URI guard into the shared one lib/agent-uri-guard.mjs was a second module over lib/uri-guard.mjs, holding one more hint table and nothing else, and the three harnesses that imported it each carried their own stdin read, entrypoint check and deny envelope — thirty-odd lines apiece for the choice between two envelope shapes. evaluateUriGuard is that module's body with the two things a host actually differs on made arguments: the hint table, and the tool names that host guards at all — Claude Code and opencode never see the shell in their matchers, pi does, and the set was previously implied by whichever table a host happened to import. The deny envelopes and the hook plumbing move in beside it, so cursor, trae and zcode keep only the envelope they answer with. * refactor(plugins): let the last four URI guards call the shared evaluator Claude Code, opencode, dsh and pi each kept a private copy of the guard loop — normalize the tool name, look it up in a local table, sweep the arguments for a viking:// URI, format the message — because their tables name different replacement tools. The loop is the same everywhere; only the table is host data, and it has to stay host data: read(uris="..."), openviking_read(uris=[...]) and viking_read(uri=..., level=...) cannot be spelled from one set of strings. So each of the four now hands its table to evaluateUriGuard and keeps only the decision its host answers with. Claude Code's table was character-identical to the shared one, so it passes the guarded set instead — read, glob and grep, leaving Bash to the model as its test has always pinned — and its stdin read, entrypoint check and deny envelope go the way the other hook guards' did. * refactor(plugins): run both doctors from one shared entrypoint The Claude Code and Codex doctors were forked from one commit and still carry the same run: parseArgs and tryJson byte for byte, expandHome inlined, the same eight sections in the same order, the same --json envelope and exit code. The copies had drifted apart in the details — Codex never learned to say which node PATH resolves to when it differs from the one running the script. runDoctor owns that run now. A harness hands it its name, its CLI, the three sections only it can produce (install, config, activity) and the spelling it uses for the configured account and user; it gets the argument parsing, the section order, the envelope and the exit code back. Codex keeps assessHooksFeature, parseFeaturesList and the isDirectRun guard its test imports, and its auth-mode verdict arrives through onSummary. Both wrappers also stop scanning shell rc files for the pre-2026-07 wrapper blocks: the installer that wrote them is gone, so only the OPENVIKING_* exports those files may still carry are worth reporting. * refactor(plugins): compose the doctor configuration section from shared segments Both doctors printed the same section from two copies that had drifted, so a check written for one host was invisible on the other. The section is now six shared segments a host calls in order, and the three checks that only one of them had — Codex's hook-budget warning, Claude Code's extra_headers X-API-Key and NODE_TLS_REJECT_UNAUTHORIZED warnings — run on both. * refactor(plugins): declare which recall knobs the request omits by default recall-core sends limit, max_tokens, query_expansion and rewrite_max_bullets only when a layer actually set them, so the server's own defaults stand where the user expressed no preference. That decision lived in five hand-written `<name>Configured` projections and nowhere in the schema, so a loader had no way to know which knobs it owed a flag. The schema marks the four now, and a test ties the marked set to the fields recall-core actually gates. * refactor(plugins): assemble every harness configuration in one place Six loaders spelled out the same sequence — read the credential files, resolve the knobs, resolve the peer, derive the timeouts, the log path, the user agent and the send-only-when-configured flags — and each spelled a slightly different subset of it. ov.conf's `<harness>.authMode` reached only the three that passed a legacy layer, two of the four knobs the server defaults for itself were reported as configured on two harnesses, and Claude Code resolved its api key on a chain of its own. buildPluginConfig() is that sequence, once. What is left in a loader is what only that harness knows: Claude Code's four-valued credential source, Codex's on/off reading of the digest switch, opencode's three section knobs, pi's older debug-log variable, dsh's cordis input. The api key now resolves the same way everywhere, with ovcli.conf's `plugin` section ranked where the file it lives in ranks — under that file's own `api_key`, over ov.conf. Codex read it last, behind `server.root_api_key`; its doctor, its reference table and the capability reference said so, and now say what the code does. * refactor(plugins): resolve the portable bundle's connection from the shared loader The Agent Plugins package kept its own copy of the credential chain, the User-Agent, the timeout and the debug logger, so every fix to the shared resolution had to be mirrored by hand and the proxy fabricated a credential source out of whichever config file happened to load. Importing the full loader is not the answer either: it would pull the knob schema and the workspace layers into a bundle that has no hooks to run them, and those copies are committed. `buildProxyConnection()` is the connection half of that loader, which is all a stdio proxy needs. The api_key still ends at ov.conf's `server.root_api_key` even when ovcli.conf pins the chain to itself: this package ships without an installer, so an install that names only a url there has nobody to migrate its key. * refactor(plugins): run every thin-harness hook from one entry Cursor, TRAE and ZCode kept eleven shim scripts between them whose whole body was an event assignment and an import of the harness dispatcher, and each spelled the plugin root its own way — `${CURSOR_PLUGIN_ROOT}`, `__OPENVIKING_TRAE_ROOT__`, `${ZCODE_PLUGIN_ROOT}` — so the installer carried one substitution branch per spelling and a template that used the wrong one failed only at the first hook of a fresh install. The shared entry takes the event and the client from argv, keeping TRAE's trailing client-id form, and one placeholder leaves one regex to render it. The installed hooks.json is now checked against the tree it was rendered for: a command naming a script the install did not put on disk used to surface only when the host first ran it. * refactor(plugins): merge the three thin harnesses into one plugin Cursor, TRAE and ZCode were three marketplace directories running the same state machine — the same 2000ms session-start debounce, the same prompt dedup by event id and 500ms window, the same recall cache and cross-process lock — around four things that genuinely differ: the event vocabulary, the response envelope, how a prompt is read out of the payload, and how a finished turn is captured. Everything else was triplicated, so a fix landed in whichever copy the author happened to open: the URI guard existed three times over one shared evaluator, the MCP proxy three times over one shared builder, and the guard that pins how many proxies exist counted eight. `agent-hook-plugin` keeps one dispatcher on stage 4's `runHookStage`, one URI guard, one MCP proxy and one doctor, and puts the four differences in `hosts/<id>.mjs`. The adapters sit at that depth because `../../memory-plugin-shared/lib` has to resolve both here and in an installed `agent-integrations/<client>/`, so `hosts/<id>/` holds only what a host reads as configuration. The installer copies per host and reads the templates from there, and the thin harnesses get the doctor entry the full plugins have had since stage 5. The archive guard was reaching none of this: the three plugins' hooks.json named the shared entry across the plugin boundary, so nothing required their own dispatchers to be in the release archive. hooks.json names the plugin's own `scripts/hook.mjs` again, which puts every adapter back in the derived requirement set, and the checker is handed the directories the staging script declares instead of listing whatever the stage happens to hold. * fix(install): reclaim the URI-guard hook entries on uninstall `--uninstall` recognised an entry as OpenViking's only by the hook script it named, and the URI guard was not on that list. Uninstalling Cursor therefore left beforeReadFile and beforeShellExecution in ~/.cursor/hooks.json, and TRAE kept its PreToolUse entry, all of them running a uri-guard.mjs under agent-integrations/<client>/ that the same uninstall had just deleted: the host reported a failing hook on every file read and shell command until the user edited the file by hand. The uninstall filter now reclaims anything the installer wrote, by the OPENVIKING_INTEGRATION_ID the rendered command carries, the way the install-time filter and the TRAE CLI uninstall already did. * ci(plugins): select the plugin test files by glob The step named 54 paths by hand, so a new test file was only covered once someone remembered to add it there too. Seven globs select the same set: the union and the old list differ in neither direction (82 files each), and no generated shared/ copy holds a test for a glob to pick up by accident. examples/*/scripts/*.test.mjs 19 examples/*/scripts/lib/*.test.mjs 2 examples/*/servers/*.test.mjs 2 examples/*/tests/*.test.mjs 20 examples/*-plugin/*.test.mjs 11 examples/memory-plugin-shared/*.test.mjs 27 agent-plugins/*.test.mjs 1 Two things the job could not see before. A stale lib/MANIFEST or committed shared copy passed CI, because nothing looked at the tree after the generator ran; `git diff --exit-code` does now. And the two installer tests ran beside the other 80: staging the marketplace regenerates the shared copies those files import, and both fork whole installs, which is the load the two unexplained single-test failures appeared under during this work. They now run in a step of their own at --test-concurrency=1, and install-agent-hooks.test.mjs copies examples/ into a tmpdir so `--source dev` no longer installs out of the tree everything else is reading. Neither failure reproduced in five full runs here. The only failure five runs did find was deterministic and local: a gitignored openclaw-plugin/shared left behind by work that has moved to another branch, which the generator reports as stale but never deletes. The openclaw job keeps its exclude; only its counts change, measured on this branch: 42 files, 675 tests, and 2 pre-existing architecture-boundaries failures rather than 4. * ci(plugins): bring pi into the plugin version gate pi's version was a hand-written constant in config.ts, which put it out of reach of check-plugin-version-bumps.sh: the extension is copied wholesale by the installer and reports that constant as the build talking on the wire, so a shipped change under a frozen version had nothing to catch it. The extension now carries a package.json, EXTENSION_VERSION reads it through the shared readManifestVersion, and the gate watches that file the way it watches the other five manifests. The manifest changes nothing about how the extension loads or ships. pi's loader reads a package.json only for a pi.extensions field and falls back to index.ts without one; pi install parses a local path as a directory rather than an npm package, and only installs dependencies for git-cloned sources; the installer's tar copies the directory whole, so the file is in the marketplace archive and the installed copy resolves the same 0.3.0 the constant held. "type": "module" states the format node was already detecting per load, which is what the warning naming this file asked for. Against origin/main the gate still reports four plugins: pi's manifest is new on this branch, so it takes the same no-baseline skip agent-hook-plugin takes, and is enforced from the next base onward. * ci(plugins): merge the two npm plugin releases into one matrix The dsh and opencode release workflows were the same 81 lines with seven values swapped, so every fix to the publish flow had to be made twice and one copy could drift unnoticed. plugin-npm-release.yml carries the flow once and lists the two packages as matrix entries. Every way the two files differed, and how the matrix expresses it: workflow name one "Plugin npm Release"; the per-package "Publish @openviking/..." survives as the job name job name name: Publish ${{ matrix.package }} work directory defaults.run.working-directory: ${{ matrix.directory }} concurrency moved from the workflow to the job and keyed by matrix.directory, so the two packages still queue independently rather than behind each other push paths the union of both plugin directories plus this filename; the shared lib entry was already identical in both install command matrix.install: npm ci for dsh, npm install for opencode, which ships no lockfile name assertion compared against matrix.package through an env var instead of a literal inside the node script The union filter means a dsh-only push also starts the opencode leg. That leg validates and then stops at the npm view already-published check, the same check that already made a shared-lib push a no-op for whichever package was not bumped. Everything else is byte-identical to what both files ran: permissions, checkout, node 24, the validate step, and the publish step with its NPM_TOKEN-or-OIDC fallback and --provenance. actionlint reports nothing on the new file. RELEASE.md named the opencode workflow by filename, so its bullet now names the merged workflow and both packages. * refactor(plugins): give each skill copy its own sync target SKILL_TARGETS was the one list in the generator with its own shape — a skill plus an array of directories — so the delivery flag every other target carries had nowhere to live, and nothing asserted that git holds the skill copies a host installs by path. One entry per copy, `{ skill, dir, committed }`, and the test that decides which generated copies belong in git reaches skills as well as vendored modules. Every skill copy is committed and stays that way: .gitignore covers the vendored shared/ directories only, and a host installs a skill by copying its path out of this repository, so a copy git does not hold ships nothing. Nothing else moves — the same two skills reach the same six directories and the generator writes the same bytes. openclaw-plugin and agent-plugins stay out on purpose: their skills are different files over different tool surfaces, not copies of these. * refactor(install): move the installer's JavaScript into lib/install install.sh carried a 328-line JSONC editor and three near-identical hooks/mcp merges as node heredocs. Nothing could exercise them except running the whole installer against a scratch HOME, and they had already drifted: the uninstall copy of the ownership test learned to reclaim the URI-guard hook entries and the install copy never did, so reinstalling left a stale guard behind. jsonc-edit.mjs is the editor, moved verbatim behind one entry point. host-json-config.mjs is the merge, once: a single read that refuses to overwrite what it could not parse, a single atomic write, a single answer to "is this hook entry ours", and three commands over them — write, remove and merge-zcode. The ownership list is the uninstall side's, so an install now replaces a stale uri-guard.mjs entry instead of appending beside it, and the zcode fold reclaims by the same rule as the other three hosts rather than by a substring of its own. lib/install/ is the installer's code, not the plugins'. No shared module imports it, so it enters no vendoring closure and no lib/MANIFEST; one `./install/...` import from a module that hooks do import would put it in both, and install-lib-closure.test.mjs now fails on exactly that. Finding it is the one thing that is not a straight move. `--uninstall` runs before any source is resolved and the documented uninstall pipes this script from a URL, where there is no sibling directory to read — reaching for the checkout there would clone a repository just to remove hooks. So the assembled runtime keeps a copy of the directory beside the manifest modules, and install_lib_dir looks next to the running script, then there, then at the checkout. install-opencode-jsonc.test.mjs is six in-process cases over the editor — comments wherever they sit, trailing commas, a single-quoted value holding a brace and a `//`, an existing nested mcp object, idempotence, and a server the user disabled — plus the one end-to-end install that proves install.sh still hands the module the right arguments. The capability reference's install.sh line count goes with them; it was already wrong and this makes it wronger. * refactor(tests): share the mock server and the hook runner Seven test files carried their own copy of the same three helpers and the copies had drifted: two writeJson signatures, and two withMockOpenViking contracts, one of which called .catch() on the handler's return value and so could not take a synchronous handler at all. Three more carried a hook runner that turned a non-zero exit into a rejected promise — a result no test could assert on, which is why the async ZCode test spawns the hook by hand where it wants to read the exit code. testing/support.mjs already held buildConfigForTest and is the right home for the rest: it sits outside lib/, so sync.mjs vendors none of it into a shipped plugin, and a helper imported from a *.test.mjs file would have registered that file's own tests a second time. Two things the shared versions do that no copy did. The mock logs every request it saw and hands the log to the callback, which the three tests that only recorded pathnames now assert against instead of keeping an array of their own. And runHookScript never rejects: the exit code comes back like stdout does, and the call sites say expectExit where a clean exit is part of the expectation. writeJson takes the status last and defaults it to 200, so the two-argument callers read unchanged and claude-code's auto-capture moves its status to the end. claude-code's auto-recall is not one of the seven — it never carried readRequestBody — and keeps its two local copies. * refactor(tests): make the config and credential contracts shared The credential chain is one resolver every harness calls, but only codex tested it, so a change to the chain broke six harnesses and one suite reported it. That file moves to memory-plugin-shared, where the glob picks it up as the shared contract it always was. The five per-harness config suites had the same problem in reverse: each one re-tested the layer stack, the peer order and the cwd rule that plugin-config.test.mjs already holds every loader to, and buried the one or two cases only that harness answers. Three contract tests finish that file — a knob walked through every layer, a default told apart from a choice, and the workspace layer read from the cwd the loader is handed rather than the process's — and the per-harness suites keep what is theirs. Trimmed: claude-code and codex scripts/config.test.mjs, opencode and pi tests/config.test.mjs. Every case deleted from them is answered by a shared one in plugin-config.test.mjs, workspace-peer.test.mjs or the credentials contract this commit moves. dsh loses nothing: the cordis input is a layer no other harness has. Two things only a harness knows had no test at all, and now do: claude-code's tri-state digest mode, which it reads from the shared on/off switch under either env spelling, and codex's boolean reading of that same switch, where a compressor counts as configured only once it has been told what to run. * refactor(tests): test the MCP proxy contract once The proxy is one shared module, but its protocol contract was tested through two harness entrypoints — twelve cases under codex, one of them again under opencode — and both entrypoints re-exported the factory they import purely so a test could reach it. A change to the core meant editing two suites, and the opencode case differed from codex's only in which User-Agent string its fixture made up. The twelve cases and their fixture move to mcp-proxy-core.test.mjs, beside the two transport-error cases already there, and the file now has one proxy builder: it takes a fetchImpl, so an injected fetch and a real loopback upstream are the same harness. Neither host file keeps a case of its own — every one of the twelve exercises the shared core through the harness's config shape, which mcp-proxy-config.test.mjs already covers per harness. One case the split never had: a proxy with no local tool provider. Most harnesses run it that way, so the local-tool branches have to stay invisible — the listing is the upstream's, and a call named like a local tool is still the upstream's to answer. `examples/*/servers/*.test.mjs` now matches nothing. The pattern stays, for the next case that really is one host's, under nullglob so it contributes nothing rather than reaching node as a literal; opencode's test script drops it. opencode keeps its servers/mcp-proxy.mjs export map entry, which describes the entrypoint the package exposes, not the deleted test. * refactor(tests): test the shared runtime where it lives Three suites tested shared code from a harness directory, so a second harness reusing that code was covered only by accident. The recall timeout case, the queue cases that never touch cc's ov-session, and findLastHumanTurnIndex now sit beside the module they exercise. findLastHumanTurnIndex moves with its cases: it is a generic helper on the shared turn shape, and codex's capture-utils.mjs already re-exports everything shared, so its one caller needs no edit. * refactor(plugins): label each credential from the chain that resolved it Three doctors each re-walked ovcli.conf and ov.conf to say where the url, the key and the identity came from — the walk `resolveOpenVikingCredentials` had just finished. They were not the same walk. Claude Code and Codex rebuilt the key's origin by hand and ignored the `apiKeySource`/`credentialPath` every config now carries, so a key the chain took from one layer could be reported as coming from another; the thin-harness doctor read `apiKeySource` but then let ov.conf's harness block name the account under a pin the chain stops above. `credentialSources(cfg, cliConf, ovConf, { section })` lives in doctor-core now, with the section defaulting to the calling harness. The key's layer comes off `cfg.apiKeySource`; the files are asked only which field inside that layer holds the value, which is the one thing the layer cannot say — `plugin.<harness>. apiKey` from `api_key`, a harness block from `server.root_api_key`. The identity follows the chain step for step, plugin section included. The stage-5 guard could not see the three copies because the core never exported the name. It does now, so a fourth one fails that test. The two harness test copies move to doctor-core.test.mjs, beside the function they describe. * ci(plugins): generate OpenClaw shared runtime before tests Co-authored-by: TRAE CLI <traecli@bytedance.com> * docs(plugins): define hook and MCP development standard Co-authored-by: TRAE CLI <traecli@bytedance.com> * docs(plugins): publish bilingual development guide Co-authored-by: TRAE CLI <traecli@bytedance.com> * fix(plugins): restore runtime and distribution contracts Generate shared dependencies before validation and source installs. Keep Codex retries owned by the transcript cursor and defer threshold commits until the tail is delivered. Enforce the shared hook enable gate, restore auth-mode environment precedence, and preserve pi request options. Co-authored-by: TRAE CLI <traecli@bytedance.com> * refactor(plugins): use a host-neutral hook manifest Move the shared hook integration metadata out of the Claude-specific directory and update version checks, diagnostics, installation verification, archive validation, tests, and documentation. Co-authored-by: TRAE CLI <traecli@bytedance.com> --------- Co-authored-by: TRAE CLI <traecli@bytedance.com>	2026-09-15 18:28:20 +08:00
t0saki	1d89f8d465	feat(plugins): derive the workspace peer from git, and let a repository carry its own config (#4595) ... * fix(plugins): stop dropping ovcli.conf's plugin section, and unrot the sync test `ov config add|edit`, the config wizard and `ov config switch` all rebuild ovcli.conf from the `Config` struct, which has no `plugin` field and no catch-all — so every write silently deleted the whole `plugin` section the memory plugins own. `write_config_file` now carries over the top-level keys `Config` does not model, `save_edited_config` reads them from the old name on a rename, and `activate_config` keeps the active file's. Modeled keys are deliberately not carried over: one that is `None` was cleared on purpose. `KNOWN_CONFIG_KEYS` decides what counts as modeled, guarded by a test that fails when a struct field is added without listing it. sync.test.mjs kept its own copies of the target lists and they had drifted — dsh, opencode and agent-plugins were each missing modules the sync ships, so a stale vendored file passed CI. It now imports the lists from sync.mjs (whose `main()` moved behind an entrypoint guard) and additionally fails on a module no target claims or a banner-carrying orphan no target lists. Also in this hygiene pass: - postRecall dropped `peer_scope` on any 400/422, silently widening recall from the caller's own peer to the whole user root. It now retries only on an unknown-field rejection, remembers the downgrade so every turn stops paying for a rejected request, and both doctors warn while that memo is live. - The session peer pins (Claude Code's `ws-peer-*.json`, Codex's `workspacePeerId`) carry a version, so a pin written under one derivation rule cannot outlive it. Derivation is unchanged, so today they re-derive to the same value. - recall-session-wiring.test.mjs was never registered in CI and had rotted against a fourth RecallManager argument; assertion fixed and registered. * feat(plugins): layered workspace configuration A workspace can now carry `<root>/.openviking/config.json`, which a team commits, and `config.local.json`, which stays private; a per-machine registry under `~/.openviking/workspaces/` sits above both so the user keeps the last word over any repository they clone. All three share one schema and one merge, and they slot in exactly where ovcli.conf's `plugin` section already does, so `OPENVIKING_*` still wins over everything. Three modules, synced to all seven plugin targets: - workspace-identity.mjs finds the workspace root and reads git's own idea of what the repository is called, using only filesystem reads. No `git` subprocess: hooks are fresh Node processes on prompt-level paths with budgets as tight as Codex's 3s SessionEnd, and this keeps working where git is absent from PATH or would refuse the repo over dubious ownership. Worktrees converge through `commondir`; submodules stay separate; `$HOME` and `/` are never workspace roots. - workspace-config.mjs discovers, parses, filters and merges the layers, and records per-key provenance — which layer won and what it covered up. - workspace-registry.mjs keeps one file per workspace rather than one listing them all, so concurrent hooks cannot lose each other's writes, and treats a path whose git identity has changed as a miss rather than inheriting the previous repository's peer. These files are trusted without a prompt, because a hook is non-interactive and any approval gate degrades into "run one command per workspace first". What is refused instead is structural and costs nobody anything: connection and credential keys are stripped loudly, `${VAR}` is never expanded, and `cli_config_profile` — which decides which credentials reach which server — is registry-only and name-only. What a committed file switches off is announced in doctor rather than blocked. An adversarial review pass over these three modules found 18 defects, all fixed here and each now covered by a test. The one that mattered: `JSON.parse` keeps `__proto__` as an own property, so a 128-byte committed file could write straight into `Object.prototype`, and since `process.env` reads through the prototype chain and the environment outranks ovcli.conf, that set `OPENVIKING_URL` and `OPENVIKING_API_KEY` for the whole process — silently shipping the user's real API key to an attacker's host. Prototype keys are now dropped with a warning in both the strip and the merge. The rest: unbounded recursion (a 4KB file could take out every sibling layer), the identity cache storing a remote's embedded token at 0644, worktrees under a directory named `modules` misread as submodules, the registry's negative-evidence check being inert in its only caller, `Number(null)` pinning cost knobs to a bound, and provenance lying when two layers disagree about a key's type. `.gitignore` no longer ignores all of `.openviking/`, which would have stopped a team's config.json from ever being committed; doctor warns when a workspace still does. The schema maps `capture.commit_token_threshold`, matching the knob the loaders actually read — the RFC's example named a turn-based one that does not exist. * feat(plugins): derive the workspace peer from git, configurably The peer a workspace writes its memories under was the working directory with every non-alphanumeric byte turned into a dash. That made the identity an accident of where the repository happened to sit: a clone on another machine, a rename, a worktree, or simply `cd examples/` each minted a separate, empty namespace, and there is no server-side rename or merge to recover from it. The default is now git's own idea of the repository. `peer.source` decides the rule and reads from every layer — `OPENVIKING_PEER_SOURCE`, ovcli.conf's `plugin.peerSource`, or `peer.source` in a workspace file: - `git` (new default) ≡ `["{git_remote}", "{git_root}", "{cwd}"]` — the normalized origin, else the repository root, else the working directory. No preset adds a prefix; a path-derived id already starts with `-` on POSIX, so it cannot collide with a remote-derived one. - `cwd` — the old rule, byte for byte. - `none` — send no peer. `OPENVIKING_WORKSPACE_PEER=0` still means this. - Any template, or a list tried in order, over `{git_remote}` `{git_root}` `{cwd}` `{dir}`. Substitution is all-or-nothing: an empty variable falls through to the next template rather than leaving a half-formed shared id. So `/Users/x/Dev/OpenViking/examples/codex-memory-plugin` with origin `git@github.com:volcengine/OpenViking.git` is `github.com-volcengine-openviking` from any subdirectory, worktree, machine or clone. Every clone of one repository shares one peer; a fork has a different origin and stays separate, and `gh pr checkout` of someone else's PR does not change origin, so reviewing does not move a session's memory. Nobody has to migrate. The pre-git id is always recomputable locally, so `resolveEffectivePeerId` returns it alongside the effective one and recall still reaches it: under the default `peer_scope: "all"` the server's cross-peer sweep already covers it for free, and under `"actor"` — where that sweep is off by definition — the plugin asks that peer separately, as itself, which is cheaper and reaches more than a bare cross-peer read would. There is no deadline on this. Wired through all five recall paths; doctor names the previous peer and says which of the two is carrying it. `source` keeps its three values because five call sites compare it against the literal `"workspace"` to decide whether a session pin may be reused; the new `origin` field names the template that actually produced the id, and doctor prints it. Both session pins bump their version, so a pin frozen under the old rule cannot outlive it. * docs: the workspace peer comes from git, and a workspace can carry config Every page that described the peer as the working directory with its non-alphanumerics dashed now describes `peer.source` and the git default, with the presets, the template variables, the clone-vs-fork identity semantics, and why no migration is required. The capability reference gains the three new shared modules; the client configuration page gains a Workspace Configuration section covering the two workspace files, the per-machine registry, the precedence table, the v1 schema and what a workspace file may not set; the Claude Code and Codex integration pages, which never mentioned peer derivation at all, each gain a short section. All zh mirrors follow. `examples/schemas/workspace-config-v1.json` is the schema the `$schema` key in a workspace file points at, and `examples/workspace-config.example.json` is a file to copy. `ovcli.conf.example` shows `plugin.peerSource`. Two facts worth stating plainly, both verified against the loaders rather than assumed: `OPENVIKING_PEER_SOURCE` and the workspace-file layer are read only by the Claude Code and Codex plugins today, so the other harnesses run on the default and their pages document the config key rather than an env var that would be inert; and pi and dsh compute their legacy id from the process cwd, so their pages promise dual-read only under the default `peer_scope: "all"`. * fix(mcp): stop the proxies from guessing a peer out of their launch directory Three MCP proxies keyed the actor peer off `process.cwd()`, which for a long-lived server started from a static MCP config is the directory the harness happened to launch from — often the plugin's own. The Codex proxy already refused this and had a test forbidding it; the rule now holds for all of them through the same `resolveMcpActorPeerId`. The plan called for the parent process to inject `OPENVIKING_PEER_ID` at launch instead, following dsh's `mcp.mjs:27`. That only works for dsh: Claude Code and agent-plugins are launched from a static `.mcp.json`/`mcp.json` with no environment block, and OpenCode's `createOpenVikingMcpConfig` builds a command and args with nowhere to put one. So the fix is to stop guessing rather than to guess better — a proxy sends no actor peer, which is broad recall, the default. `resolveMcpActorPeerId` now warns and widens where it used to throw. Refusing to start took away every memory tool because a scope preference could not be honoured, which costs the user far more than the wider search does; the warning says which two settings would scope it. * test(pi): follow the git-derived peer default rather than pinning the cwd id * fix(plugins): warn on the camelCase spelling of a connection key too The projection into harness knobs is an allowlist, so `apiKey` in a workspace file could never take effect — but it vanished without a word, which reads as acceptance. It is refused by name now, like its snake_case twin. * feat(cli): ov workspace show, and ov peer link|migrate|forget-previous `ov workspace show` answers "which layer actually set this" the way `git config --show-origin --show-scope` does: the workspace root and how it was found, the git remote, every template variable, the effective peer and the template that produced it, each config layer with whether it applied, and per key the effective value plus everything it shadowed. That question matters here because three languages read this configuration and each could drift. So the Rust reader is not a paraphrase of the JS one — the two were run side by side over the identity helpers, the merge with full provenance trees, the file-read rules and the registry's raw bytes, and made byte-identical. That comparison paid for itself: it caught `serde_json::Map::remove` being a swap remove under `preserve_order`, which reshuffled a registry file the JS half reads on every rewrite. It also caught the divergence that would have made the command a liar. ovcli.conf's `plugin` section speaks the flat knob names a harness loader reads (`peerSource`, `recallLimit`); a workspace file spells the same settings nested (`peer.source`, `recall.max_items`). Both are one chain in `loadPluginSettings`, and the port had merged the flat file into the nested tree, so `plugin.peerSource: "cwd"` in ovcli.conf left `ov workspace show` reporting the git-derived peer while every plugin sent the cwd-derived one. `ov peer link <id>` pins a peer for this workspace in the registry — the way out of a fork that should share the upstream's memory, or a legacy id worth keeping. `ov peer migrate` moves a peer's memories and resources with the fs mv API, reporting the plan by default and requiring `--apply`; the server has no merge semantics, so a collision is refused with the colliding path rather than overwritten, and a listing that fills its limit aborts rather than planning from a truncated view that could hide one. `ov peer forget-previous` clears the recorded ids. `workspace show`, `peer link` and `peer forget-previous` are local and do not require ovcli.conf; `peer migrate` talks to the server and does. Both config gates and the hand-rendered help are registered, with a test pinning the gates against each other. A test now reads FORBIDDEN_KEYS, REGISTRY_ONLY_KEYS and FREE_FORM_SECTIONS out of the JS module and compares them to the Rust constants, because those lists are what someone fixing a bug in one language edits — and they had already drifted once while this was being written. * build(cli): record the sha2 dependency edge in Cargo.lock Already vendored for other workspace members; ov_cli now uses it for the registry slot hash. * test(plugins): the fixtures the plan named that were still missing A moved or renamed repository keeping its identity is the change's whole point and had no test; a shallow clone was worth pinning because it is exactly what the rejected root-commit scheme could not answer; and the registry's read-modify-write window between two hooks of one session is now written down as a test rather than only as a comment. * fix: the defects a plan review turned up An independent review against the plan this branch was built from found ten real defects. Each was reproduced before being fixed and is now covered by a test. The four that broke a promise the feature makes: - The workspace config layer was resolved from the hook process's own working directory, not from the `cwd` on its stdin payload — which is the authoritative one. A hook started in one repository while the session sits in another applied the wrong `.openviking/config.json`: its peer, its bypass patterns, its `capture.enabled`. `loadConfig` now takes the directory, and every hook that receives one re-resolves with it. Late re-resolution is safe precisely because connection and credential keys are structurally forbidden in a workspace file, so `baseUrl` and `apiKey` cannot move under an already built client — the gates that a workspace can switch off moved below the parse so they are decided on the right config too. - Codex threw away a workspace file's `peer.id`: it returned the credential chain's peer verbatim, so `{"peer": {"id": "team-a"}}` did nothing. Claude Code had always honoured it. - Claude Code's session pin returned only the id and source, dropping `legacyPeerId` — so from the second hook of a session onward, dual-read stopped asking the peer that holds everything written before the derivation changed. Silently, and exactly where it mattered. - `ov peer migrate` read the source peer with the actor-peer header set, which the server refuses for another peer's path, and treated every `stat` error as "does not exist". The common case — an `actor_peer_id` in ovcli.conf — got a cheerful "Nothing to migrate" instead of a 403. It now uses a client with no actor peer and tells a real error apart from an empty source. The rest: - A directory outside any repository is a workspace again. It had no root at all, so a `.openviking/config.json` there was ignored entirely. `$HOME` and `/` are still never roots, now judged on the starting directory rather than on where the upward walk stops, and `git_root` stays empty outside a repository so the `git` preset still falls through to `{cwd}`. - The registry slot is keyed on identity, not path. Two linked worktrees of one repository are one workspace — one peer, so one set of settings and one `ov peer link` — and keying on the checkout path split them in two. This also makes crossing two repositories physically impossible rather than merely detected. (Their `config.json` files still follow each checkout; those are files on a branch.) - git folds section and key names to lower case, so `[Remote "origin"]` with `URL = …` is a remote `git config` reads and we did not. A quoted subsection stays case-sensitive. - `min_client_version` warns instead of being silently kept as data, and still never blocks. - `cli_config_profile` was validated and then never used. It now selects `~/.openviking/ovcli.conf.<name>` before credentials resolve — registry-only, name-only, and a hard error when the profile is missing, because quietly authenticating somewhere the user did not choose is the failure the key exists to prevent. - `ov workspace show` is exempt from the language gate. It is a diagnostic and has to work on a machine where no language was ever chosen; `peer link`, `migrate` and `forget-previous` mutate state and still gate. - `ov peer link` records the peer it replaced, so a later `migrate` with no `--from` finds it instead of falling back to a recomputed cwd id. Two more the plan asked for that were missing: doctor now checks the knobs *inside* ovcli.conf's `plugin` section — it was on the allowlist, so until now `peerSorce` sat there doing nothing with no complaint — and suggests the key you probably meant. A test derives the known-knob set from what the two loaders actually read, so the list cannot rot into one that rejects a real knob; it caught a missing entry the moment it was written. The RFC is archived at docs/design/, with the three claims implementation disproved corrected in place: the knob is `commit_token_threshold`, `__self` and `ext-` are not reserved server-side, and a worktree converges its identity rather than its config file. * revert(cli): withdraw ov workspace and ov peer from this branch The command surface these two files added was larger than the feature they served: 4792 lines of Rust for `ov workspace show` and `ov peer link|migrate|forget-previous`, against a change whose whole point is what the hooks send. None of it had reached a user-facing document — only the RFC named it — so it goes back out whole and the branch becomes a plugin change plus one CLI bug fix. Restored from the branch's merge-base rather than from origin/main, since main has moved on since the branch was cut and those commits are not this branch's to carry. `sha2` was pulled in only by `workspace.rs`, so its dependency edge leaves with it. What stays is `config.rs` and `config_wizard/store.rs`: `ov config add|edit` dropped the whole `plugin` section because the wizard round-tripped the file through a typed struct, and that fix has nothing to do with the withdrawn commands. The registry under `~/.openviking/workspaces/` stays too, as a layer the plugins read. Nothing writes it for now; `ov-memory-doctor` prints the path it expects, and the file is small enough to create by hand. A writer can come back on its own merits. * fix(plugins): derive a peer only inside a git repository Codex desktop opens a directory per task — `~/Documents/Codex/<date>/<slug>/` — and none of them is a repository. The `git` preset ended its fallback chain at `{cwd}`, so every one-off task minted its own empty peer, and each new one started with no memory. Nine such directories here, nine peers. There is nothing app-specific to read: the state file that lists those threads is Electron-private, a megabyte wide, desktop-only, and would have to be parsed inside SessionEnd's 3-second budget. The signal that generalizes is structural — the directory is not a repository, and nothing in it says it is a project. So the default chain is now `["{git_remote}", "{git_root}"]` and stops there. A directory that is neither a repository nor marked gets no peer at all, and what is remembered in it goes to the user-level space, which is where it went before peers existed. Deriving an identity from a bare path is what `peer.source: "cwd"` is for, and it is a word away. Naming such a directory is the other half. `findWorkspaceRoot` now also stops at a directory holding `.openviking/config.json` or `config.local.json`, so a marker file works from any depth below it, the way a repository does — and when that marker sits inside a repository the git variables still resolve to the enclosing repository, so marking a subdirectory of a monorepo does not split the default peer. `{git_root}` is the repository's root, `{dir}` the workspace root's name whichever made it one. Nothing moves. When no template resolves, the pre-git id is still computed and returned as `legacyPeerId`, so `peer_scope: "actor"` keeps asking for it and `"all"` keeps sweeping it. Two doctor bugs fell out of the same walk: `checkWorkspace` read `git.kind` unconditionally and threw wherever there was no repository, and the peer block warned "set peer.source to git" at a directory where `git` is exactly what is already set and correctly resolves to nothing. It now says why no peer is sent, and prints the file to create. * docs(plugins): say how to give a directory its own peer, to users and to agents The behavior change is only useful if the reader can act on it, and two kinds of reader have to: the person whose scratch folder stopped having a memory, and the coding agent they ask about it. `docs/{en,zh}/configuration/02-client.md` is the one place that spells the rule out, and everything else links to it. It gains "Give a Directory Its Own Peer", which opens with the file to create and then the ladder above and below it; "By Situation", eight rows from fork to throwaway folder; and "Recall Isolation", which separates where memories are written from what is read back, names the server's per-category penalties, and states the cost of sending no peer outside a repository — a user-level memory is read at full weight in every project afterwards. The eight integration pages, both capability references, six plugin READMEs, the changelog and the schema stop promising a fallback to the working directory. Checking those claims against the loaders turned up one that was never true: opencode, dsh and pi do not read workspace files at all, so a `peer.id` written for them does nothing. Said plainly rather than left to be discovered. For agents, `openviking-memory/SKILL.md` gains ten lines on where memories are filed — it is the skill that fires when someone asks why a folder has no project memory, and it had nothing to say — and both `ov-memory-doctor` references gain a table from what the user says to the exact key to write. `ov-memory-doctor` prints the same snippet, so an agent that runs it needs no further reading. One snippet has to be identical in twenty places for any of this to hold, so `WORKSPACE_PEER_HINT` is a constant the report builds its line from, and `peer-guidance.test.mjs` asserts it appears verbatim wherever it is promised, that no page still spells the retired chain or names a command this branch withdrew, and that every variable the canonical page documents is one the code substitutes. It asserts no prose: rewording a page must not turn it red. * fix(plugins): reject an unrecognized peer.source instead of using it as an id `peer.source` accepts a preset name, a template, or a list of templates, and anything that is not a preset was treated as a template. A template with no `{...}` in it renders to itself, so a typo became the peer: `"Git"` wrote every memory under a peer literally named `Git`, and `"gti"` under `gti`. Silently — the wrong namespace is indistinguishable from an empty one until someone notices their project has no memory. A bare string that is neither a preset nor contains `{` now warns and falls back to the `git` default. A list is still taken at face value: writing one is explicit enough that a typo inside it is a different kind of mistake. The warning travels through an optional `onWarn` callback falling back to stderr, matching `resolveMcpActorPeerId` in `mcp-proxy-config.mjs` — there is no warnings array in reach, because `resolveEffectivePeerId` is called from the hook runtime and from four harness config loaders, none of which thread one. * docs(plugins): say what each harness can actually do with a peer The peer documentation promised the same thing everywhere, but only the Claude Code and Codex plugins read workspace configuration files. `loadPluginSettings` is called from exactly two loaders; the other harnesses build their config from their own file plus the environment. So a reader following the docs under pi, dsh, opencode or cursor would create `.openviking/config.json` and watch it do nothing. The skill is the sharpest case: `openviking-memory/SKILL.md` is synced to cursor and dsh as well, and it told an agent to write that file. An agent would have done it, reported success, and changed nothing. It now names the two harnesses that read it and points everyone else at `OPENVIKING_PEER_ID`. The integration pages had started teaching the recipe and then retracting it in the same sentence, which is worse than not mentioning it; they now carry the one instruction that works there, and link to the canonical section for the rest. The capability reference gains the same qualification, next to the paragraph that already says only two harnesses read those layers. Two smaller corrections. The doctor references had the same question answered twice, once in the peer table and once in the troubleshooting table 140 lines below; the troubleshooting row survives, since it carries a diagnostic column. The RFC still archived implementation notes for the CLI this branch withdrew, which would read as a description of commands that exist. `peer-guidance.test.mjs` guards this alignment, and had two flaws of its own: it swept the changelogs, which are generated from GitHub Releases and would go red on a release note nobody wrote by hand, and it sliced a page between two headings with `indexOf` without checking either was found — renaming the closing heading would have silently scanned to end of file. Also here, because it is the same kind of mismatch: the zcode MCP proxy sent an actor peer under broad recall, where the other proxies leave the header unset. It now routes through `resolveMcpActorPeerId` like they do. The dsh proxy deliberately does not — its parent process resolves the peer per session and injects it into the child environment, so it is not guessing at a launch directory, and that reason is now recorded next to the line. * fix(plugins): ship and install only the shared modules a plugin imports Three new modules were fanned out to all seven plugin directories in one hunk, and only two plugins call them. That left dead weight in five directories, and it broke three installs. The install is the part that mattered. `install.sh` copies a hand-written list of shared files into `~/.openviking/agent-integrations/memory-plugin-shared/lib`, where cursor, TRAE and TRAE CLI import from. The list names `workspace-peer.mjs` but not `workspace-identity.mjs`, which `workspace-peer.mjs` imports — nor `workspace-config.mjs`, which identity had come to import for three filename constants. Copying exactly that list and importing the hook runtime fails with ERR_MODULE_NOT_FOUND, so every hook of those three harnesses would have died on startup. Nothing tested the list. `CONFIG_DIR_NAME`, `TEAM_FILE` and `LOCAL_FILE` now live in `workspace-identity.mjs`, which is where the walk that recognises a marked directory needs them; `workspace-config.mjs` imports and re-exports them, so no call site moves. Identity has no library-internal dependency left, which is what makes the installed set closed at fifteen files instead of pulling the whole configuration layer along behind it. `install-lib-closure.test.mjs` derives both sides — the list parsed out of the shell script, and the transitive imports of the three entrypoints — and fails in either direction, so neither a new dependency nor a stale entry can go unnoticed again. With identity standing alone, the fan-out can follow what is actually imported. `sync.mjs` moves from arrays chained by spread — where the harness that does not need a file is often the one the array is named after — to explicit per-target lists. `plugin-config.mjs`, `workspace-config.mjs` and `workspace-registry.mjs` leave dsh, pi, opencode and zcode, none of which import them; all four workspace modules leave agent-plugins, whose only importer was deleted a half hour after they arrived and whose peer is environment-only by design. That is about 4500 lines of vendored code that said something the code did not do. The registry loses its write path in the same spirit. `writeEntry`, `rememberPreviousPeer` and `listEntries` had no caller outside their own tests: the CLI that would have written them was withdrawn from this branch. `readEntry` and `entryPath` stay, because a hand-created entry is still read and the doctor still prints where to put one. `cli_config_profile` goes with them — the whole mechanism, down to the documentation that described it, since nothing ever resolved a profile through it. This is not a new policy. `HARNESS_KEYS` already carried the rule in a comment, added the day after the same speculative fan-out happened in August: add a key as its loader starts calling `loadPluginSettings`, not before, so the section never promises a knob that silently does nothing. * fix(dsh): thread peerSource into the per-session peer The integration page documents `peerSource` in dsh's Cordis patch, but `stateFor` never passed it to `resolveEffectivePeerId`, so the key resolved to nothing and every dsh session ran on the default derivation. Pass it, and keep the pre-git id alongside so dual-read reaches memories written before the default changed. * docs(plugins): correct the shared-layer counts after the distribution changed The capability reference still described the pre-branch distribution: 18 library modules against 23, per-target counts from before each target stopped receiving the workspace configuration layer, and `workspace-config` / `workspace-registry` listed as reaching every JS harness when only claude-code and codex load them. It also still said `cli_config_profile` was registry-only, and that the registry is written for you. * docs(rfc): lead with a TL;DR of the workspace config and peer source proposal * feat(plugins): let peer.source name the harness with {harness} The peer templates could describe where a checkout sits but never which agent was running in it, so one repository could not keep a separate memory per agent even when its user wanted that. The harness name was already in every config, only baked into the User-Agent string. No preset uses the new variable: sharing one project memory across agents is the more useful default, so splitting stays opt-in via a template such as "{git_remote}-{harness}". It is composed at render time rather than in the workspace identity, whose result is cached under a cwd-only key that two harnesses in one directory would share. * docs(rfc): record git_branch and peer.command as directions, not deliverables * chore(plugins): resync the openclaw vendored recall-core after the rebase * test(opencode): follow resolveEffectivePeerId's widened return shape Also mark the openclaw shared copies generated, the way every other sync target already is.	2026-09-04 12:50:43 +08:00
t0saki	7200cdb176	feat(codex): commit on SessionEnd hook, keep SessionStart sweep as fallback (#4429) ... * feat(codex): commit on SessionEnd hook, keep SessionStart sweep as fallback Codex ships a SessionEnd hook since rust-v0.145.0. The plugin now commits the OpenViking session from that hook (marker + detached worker within the 1s/3s budget, catching up turns the last Stop never sent) and reduces the SessionStart active-window heuristic to an ended/idle sweep. Adds a per-session lock so the Stop worker, PreCompact, SessionEnd worker and sweep no longer clobber each other's state writes, stops an unreadable transcript from resetting the capture cursor, and merges the three copies of the HTTP/transcript helpers into scripts/ov-session.mjs. * fix(codex): guard SessionEnd partial catch-up, make the sweep catch up, token the end marker, own the lock Review follow-up for #4429: SessionEnd no longer commits after an incomplete catch-up (tail turns were archived away); capture hooks record transcriptPath so the SessionStart sweep catches up before committing; the .ended marker's timestamp acts as a token that the SessionEnd worker and the sweep re-verify under the lock, and Stop/PreCompact/resume only clear markers older than their own start; withSessionLock stamps an owner file and takes over stale locks by atomic rename with an inode check so a taker never moves a lock a racer just created. Plugin 0.8.1. * fix(codex): catch up ended sessions with no live id, guard unreadable transcripts, make the end marker and the lock takeover race-free Four defects found reviewing the SessionEnd commit path: - The SessionStart sweep cleared the `.ended` marker of a state with no live ovSessionId before taking the lock, so a session PreCompact had released and that then produced more turns lost its tail when its SessionEnd worker died. A marker now always enters the lock, and only a catch-up that finds nothing new may clear it. - catchUpTurns reported an unreadable transcript as an empty one, so all three callers committed and archived a session whose tail they never saw. It now reports `unreadable` and they keep the session live for a later retry. - clearEnded read the marker's timestamp and then removed the path, deleting a marker a concurrent SessionEnd had just written. The timestamp moved into the filename, so a conditional removal targets an immutable path. - session-state.test.mjs was missing from the CI test list. Also replaces the lock's stale takeover: renaming the directory aside left the lock path momentarily absent, which let another racer's mkdir succeed next to the taker. Takers now race for the `owner` file inside the directory instead. * fix(codex): make the end marker's generation unique within a millisecond Date.now() alone is not a generation: two SessionEnd hooks in the same millisecond produced the same marker path, so the first one's conditional removal took the second one's marker with it. The marker is now created exclusively and its timestamp bumped until that succeeds; bumping rather than randomizing keeps the names ordered, which is what the `before` cutoff compares. The race test drops its artificial 2 ms gap and runs 500 iterations.	2026-08-31 12:23:33 +08:00
t0sakiand7487	a61c57bf59	fix(codex): preserve transcript cursors across commit and compaction (#4191) ... * fix(#4058): [Bug]: Codex memory plugin replays historical turns after resume or transcript compaction Fixes #4058 Ref: https://github.com/volcengine/OpenViking/issues/4058 * fix(codex): retire committed cursors and keep activity-based concurrency Preserving the transcript cursor after a commit stops the replay, but it also means nothing deletes state files any more: clearState() lost its last caller, so every codex session — including ones that never captured a turn — leaves a file behind, and listStates() reads all of them on every SessionStart. The sweep now retires cursor-only states in the same pass: a real cursor is kept for resume until OPENVIKING_CODEX_COMMITTED_TTL_MS (default 30 days, past the life of the codex rollout it indexes), and a state that never captured anything goes on the idle schedule, which is what the old sweep did with it. Releasing ovSessionId also wrote lastUpdatedAt, making a committed session look freshly active; saveState() takes touch:false so the field keeps meaning "last transcript activity" for both the active window and retention. Requiring a live ovSessionId to count as recently-active made the heuristic miss sessions PreCompact had just committed, which can still be running: the count is back on activity alone, and only a state with a live session is committed. Also name the shrink predicate: role === "user" covers tool results too (normalizeCaptureRole maps them onto the user role), so findLastHumanTurnIndex requires a text part, and the no-human-turn fallback to a full replay is now visible in the log instead of silent. --------- Co-authored-by: 7487 <1042653432@qq.com>	2026-08-26 11:44:40 +08:00
t0saki	a83b81715b	feat(uri)!: remove uid-less current-user shorthand in favor of viking://~ (#4196) ... * feat(uri)!: reject uid-less current-user shorthand in favor of viking://~ viking://user/<segment> (memories/resources/skills/peers/privacy/sessions without a user id) was ambiguous with a user literally named after the segment, and a user actually named e.g. "memories" was unreachable for USER/ADMIN callers. Now that the viking://~ home alias (#4167) covers the same need unambiguously, the shorthand fails closed at the request boundary instead of expanding: - resolve_current_user_uri raises NamespaceShapeError with a corrective hint naming both viking://~/<rest> and the explicit-uid form. Silently parsing the reserved segment as a peer user id would misdirect reads and writes, so rejection is the only safe removal. - Bare viking://user falls through to the canonical parser and keeps container semantics (a user key listing it sees only its own space). - The self-id escape stays: a caller whose user_id equals a reserved name keeps viking://user/<own-id> as their canonical root. ROOT-role literal parsing and the legacy viking://session alias are unchanged. - AddTargetsConfig normalizes stored legacy config spellings (viking://user/resources|skills) to the viking://~ form at validation so existing ov.conf/user_config deployments keep working; the accepted per-user spelling is now viking://~/resources and viking://~/skills. - usage_reporter keeps canonicalizing the historical shorthand found in old transcripts and additionally recognizes viking://~/memories/. BREAKING CHANGE: requests using the uid-less viking://user/<segment> spelling now fail with 400; use viking://~/<segment> or an explicit viking://user/{user_id}/<segment> URI. * refactor(clients): migrate first-party emitters to the viking://~ home alias Every in-repo client that emitted the removed uid-less current-user shorthand now sends viking://~/... instead: vikingbot fallbacks and default sentinels, the LangChain store/tools defaults, the shared recall-core.mjs (all synced plugin copies), the codex/claude-code/ openclaw/openwebui/dsh/zcode/pi plugin emitters, quick-app examples, Go SDK example, tau2 benchmark targets, and the eval golden dataset. Compat kept where legacy strings live in stored user configs: bot and ov_dream sentinels accept both spellings while emitting only ~, and recall-core still rewrites legacy viking://user/<reserved> config values client-side. langchain_openviking._uri now classifies viking://~ with the explicit-user shape so canonicalized server responses keep matching a ~ root. Plugin READMEs note the server requirement for the alias. * docs: replace current-user shorthand guidance with the viking://~ home alias Rewrite every EN/ZH doc and model-facing prompt that advertised the uid-less viking://user/<segment> spelling: URI concept catalogue, context-types/storage/extraction/retrieval/session/privacy concepts, configuration guide (with the legacy add_targets auto-normalization note), resources/skills/sessions/retrieval/admin API references, FAQ, capability reference, and the openviking-memory / ov-experience-memory / openclaw / ov-resources skills. The stale MCP viking://user/<path> dialect passage in the MCP guide is replaced by ~ guidance, and bare viking://user is documented as the container of user spaces. * test(api): migrate live API session-used tests off the removed shorthand tests/api_test/sessions sent uid-less viking://user/skills/... URIs to record_used, which the request boundary now rejects with 400 (caught by the API & CLI Integration Tests CI job; these tests need a live server and are not part of the local suites). The api_test client authenticates as an admin-role user key, so the viking://~ home alias expands for it. tests/api_test/common/test_edge_cases.py is left as is: it asserts a 400 for a non-resource add target, which still holds.	2026-08-21 19:00:19 +08:00
t0saki	33043cb1b8	feat(plugins): expose session commit trace IDs (#3977) ... Preserve result.trace_id across plugin HTTP wrappers, include it in commit success and failure logs, and surface it in user-visible commit confirmations where supported.	2026-08-13 23:29:41 +08:00
3 people	2cc96e393e	feat(retrieval): assemble auto-recall context server-side via /search mode="context" (#3534) ... * feat(retrieval): assemble auto-recall context server-side via /search mode="context" Auto-recall assembly lived in every harness plugin: each one searched per memory type, read hits back one by one, and stitched a context block with its own budget and degradation rules. The implementations drifted, and the shared weaknesses showed up in production injections — roughly half of the entries degraded to a bare URI plus a score, character budgets distorted up to 6x on CJK text, and adjacent turns re-injected the same memories. This moves assembly into the server as one round trip. /find stays an unchanged stateless primitive. /search gains mode="context" (mode="list" is the default and byte-identical to before), and /recall becomes a thin preset over the same kernel with its v1 field names folded onto the new contract. New assembly kernel under openviking/retrieve/context_assembler/: - Token budgeting with a CJK-aware estimate replaces the character budget. - detail="auto" fills breadth-first then deepens: every candidate gets a readable floor, then overview, then full for high-scoring entries. An oversized tier falls back to the previous one instead of being truncated, bounded by max_tokens / candidates * 2 per entry. - Overview extraction dispatches by source: memory files use their leading Summary section, code files reuse code_outline signatures, long documents use a heading tree plus first paragraph. - Directory hits start at overview and read their .overview.md sidecar, since directories carry no stored abstract; their full tier stays capped at overview. v1 injected the sidecar as if it were a whole file. - Quotas generalize beyond memory types to resources and skills, with purpose presets supplying ratios when quotas are absent. - dedup_turns keeps a per-session ledger at {session_uri}/.recall_log.json so every harness inherits cross-turn dedup; exclude_uris remains as the stateless fallback. - Rendering flattens to one <memory uri=... type=... score=... detail=...> element per entry. Every tier carries its URI, so the model can always drill down through the MCP read tool. - Query expansion and digest rewriting are opt-in and fail closed: both have timeout fuses, and a failed rewrite still returns the unrewritten block. Retrieval failures are counted into stats rather than silently yielding an empty block. Plugins now send one context request, falling back to /recall and then to raw find on older deployments, and cache that outcome so only the first turn pays for the probe. The tri-state recallRewrite knob chooses between local host-CLI compression and the server digest, and client-side settings move to a plugin section in ovcli.conf. * refactor(retrieval): give context tiers a per-category default The tier ladder assumed `abstract` is a cheap summary. For memory files it is not: the memory writer stores the whole stripped body in that scalar because it doubles as the embedding text, so `abstract` costs the same as `full` and the ladder runs `uri < overview < abstract = full`. Two of the model's properties fell out of that: exempting `abstract` from the per-entry cap let a single entry eat several times the budget, and `detail` — which only ever set a ceiling — collapsed to two distinguishable behaviours across its four values, since `auto` already allowed `full` for memory. Tiers now come from a per-category constant table that treats the storage shape as a given: `events` starts at overview (the one memory type whose `# Summary` extraction is a real compression) and may deepen to full on leftover budget; every other category is served at `abstract`, which for memory already is the complete file at zero read cost and for resources and skills is the generated 256-char summary. The table carries the note to move `events` back to `abstract` once the writer stores a separate summary scalar. Falling out of that: prefetch now reads only the candidates whose planned tier needs a body rather than every candidate, `detail` becomes a real pin (start and ceiling) and additionally accepts a per-category map, and `full_score_threshold` is gone — leftover budget is spent in score order instead of behind an absolute threshold the observed score band cannot support. `auto` is still accepted on the wire as a synonym for "unset". Assembly fixes found alongside: - Removing the abstract cap exemption would turn an oversized abstract into a bare URI, so it now falls back to overview first — for memory that is a cheaper substitute, not a step up. - Rewrite timeouts were reported as failures on Python 3.10, where `asyncio.TimeoutError` is a separate class from the builtin. - `stats.rewrite_usage` read `token_tracker` off `VLMConfig`, which has no such attribute; usage was structurally always null. It now reads the model instance's tracker and reports only when the call count moved by exactly one, since that tracker is shared. - A single malformed ledger record made every deduped recall in that session fail, and the file was never rewritten, so it could not heal. Records are now coerced on read and dropped on the next write, along with records left ahead of the clock by an archive rotation. - Entries served as a bare URI no longer enter the dedup cooldown: they lost to budget pressure, not to the reader having already seen them. - The render envelope only neutralised a literal `</memory>`, so a body could forge a sibling entry with its own uri, type and score. - Flat-mode gathering re-derived the category from the URI, reading `viking://resources/backup/memories/events/log.md` as an event. - Cooled and excluded URIs are compensated with extra rows, so a fully cooled bucket falls through to the next-best hits instead of coming back empty. - `/recall` quotas overlay the v1 bucket defaults again; `{"events": 5}` had started dropping the other three buckets. - The MCP `recall` signature sent its own defaults as if the caller had, which resolved a different profile than `POST /recall`; an unknown `detail` value raised `KeyError` through the whole call instead of degrading. * feat(codex): inject profile context on session start Reuse the shared profile builder for startup, clear, and resume hooks while preserving archive injection and orphan-session status output. Co-authored-by: TRAE CLI <noreply@bytedance.com> * fix(retrieval): raise rewrite timeout default to 30s * docs(agents): document low-latency recall settings * fix(codex): prefer luna as recall compressor fallback * refactor(plugins): unify recall compression setting * feat(plugins): enable recall compression by default * docs(agents): use absolute links in image docs * fix(retrieval): address context assembly review feedback Co-authored-by: TRAE CLI <noreply@bytedance.com> * test: trim redundant context assembly coverage * fix(retrieval): address second-round context assembly review - Drop the backticked `/search` from the deprecated-recall row in both API overviews. The reference checker scans the whole row after the method cell for backticked paths, so it read the description as a route named `POST /search` and Build Docs failed on an unknown, undocumented route. - Accept ovcli.conf's full field set in both Python readers. The file's schema belongs to the Rust CLI, which writes `root_api_key`, `output`, `echo_command`, `show_progress` and `verbose` and ignores unknown keys; the two Python readers had drifted into stricter subsets, so the shipped example already failed to load in both. Adding the new `plugin` section to a working ovcli.conf would have broken `ov doctor` and every SDK client the same way. - Return 400 from `mode="context"` for a request `mode="list"` also rejects. Retrieval validates query and image_url before searching, and the gather fuse swallowed that rejection along with genuine scope failures, so a body of `{"mode":"context"}` came back 200 with an empty block instead of the documented parameter error. Runtime failures still degrade into `stats.retrieval_errors`. - Let a context request that asks for a server-side digest outlast the server's rewrite fuse. The plugin's ordinary 15s request timeout is shorter than the 30s fuse, so a rewrite that finished inside its own budget was aborted client-side, discarding the whole response — including the uncompressed block the server returns when a rewrite fails — and falling back to `/recall`. The deadline is only extended when the body actually requests a rewrite, and `OPENVIKING_RECALL_CONTEXT_TIMEOUT_MS` / `plugin.recallContextTimeoutMs` pins it. * chore(plugins): sync shared modules into the zcode snapshot * fix(retrieval): align context quotas and plugin defaults Restore cross-domain coding recall, reuse authoritative actor resource scopes, and make bucket quotas the sole width control in purpose mode. Keep plugin defaults server-owned while preserving explicit legacy limit settings through quota conversion. Co-authored-by: TRAE CLI <noreply@bytedance.com> * fix(retrieval): preserve recall compatibility Restore the deprecated recall threshold default, distinguish successful empty rewrites from compressor failures, and document legacy quota floors across coding-agent plugins. Co-authored-by: TRAE CLI <noreply@bytedance.com> --------- Co-authored-by: TRAE CLI <noreply@bytedance.com> Co-authored-by: qin-ctx <qinhaojie.exe@bytedance.com>	2026-08-05 12:31:21 +08:00
t0saki	29ba81b53e	feat(plugins): identify memory plugin traffic with a User-Agent (#3532) ... Every harness memory plugin sent OpenViking requests under Node's default `user-agent: node`, so server-side and gateway logs could not attribute traffic to a harness or a plugin version. Send `openviking-memory-<harness>/<version>` from all six harnesses (claude-code, codex, opencode, pi, cursor, trae) on both the data plane and the MCP proxy. Versions come from each harness's own manifest, or from OPENVIKING_INTEGRATION_VERSION for cursor/trae; an unreadable manifest degrades to 0.0.0 rather than throwing inside a short-lived hook. The header is purely informational — neither the open-source server nor the commercial gateway consumes it, and no auth or identity header changes.	2026-07-27 15:31:21 +08:00
t0saki	2a81edc707	Add workspace peer mode for memory plugins (#3099)	2026-07-09 17:53:36 +08:00
agentandZaynJarvis	3757a143f0	[codex] fix memory plugin recall and auth handling (#2676) ... * fix codex memory plugin backlog handling * [codex] fix install.sh validators: optional peer + parenthesized plugin list state Two false-positive validations surfaced when re-running the installer on a no-peer config after #2598 merged: 1. cached .mcp.json validator unconditionally required the X-OpenViking-Actor-Peer mapping. Since syncMcpConfig now omits that header when no peer is configured (#2598 commit 216dbb5e), the validator failed every no-peer install. Make it conditional, and symmetrically flag a stale header when no peer is configured (the same way bearer_token_env_var handles HAS_API_KEY both ways). 2. 'codex plugin list' validator grep expected '<id> installed, enabled' but codex CLI actually prints '<id> (installed, enabled)' with parens. Loosen the regex to accept both forms. - ov-credentials.mjs gains a 'has-peer-id' CLI command symmetric to 'has-api-key', so install.sh can detect peer presence via the same resolver used at render time. Underlying peerId resolution is already covered by ov-credentials.test.mjs cases 1, 2, and 3. - install.sh detects HAS_PEER_ID, mirrors the bearer pattern at the validator, and makes the codex-plugin-list regex tolerate the parenthesized state form. Verified by re-running setup-helper/install.sh non-interactively against the no-peer ovcli.conf: 'Plugin install looks valid.' now prints with no warnings. Hand-traced both HAS_PEER_ID branches of the new validator (peer-configured / peer-absent, header-present / header-absent) to confirm the four cases behave correctly. * fix codex backfill capture parsing * refactor codex background transcript capture * fix: state * fix:bug * fix: auto capture * fix * fix * Revert experimental Codex capture changes --------- Co-authored-by: ZaynJarvis <zaynjarvis@gmail.com>	2026-06-18 12:45:06 +08:00
Zayn Jarvis	c6990e4cd6	[codex] tighten memory recall, capture, and resume (#2598) ... * Fix Codex memory hook recall noise and stop timeouts * Tighten Codex recall compression output * Add Codex archive resume and capture filtering * Wrap Codex memory injection for capture filtering * Detect Codex recall compressor profile * Refresh Codex compressor profile on startup * fix codex ov credential resolution * [codex] resolve compressor profile via models_cache.json, not codex exec probe SessionStart used to spawn 'codex exec' sequentially against each candidate model to detect which one would respond — up to 3 probes × ~15s timeout each, on every session start, even on resume. The configured-on-startup default made this a guaranteed first-page-load tax of several seconds. Replace the probe with a lookup against codex's own model catalogue (~/.codex/models_cache.json, refreshed by codex CLI's etag-backed fetch). The first candidate whose slug is present wins. SessionStart now goes cache-first: load the persisted profile if any, only resolve on cache miss. The runtime compress path (auto-recall) deletes the cached profile on any compress failure so the next SessionStart re-resolves against the current catalogue. - recall-compressor-profile.mjs: * loadCodexModelsCache(env) reads ~/.codex/models_cache.json; missing cache yields {present:false,slugs:Set()}. * resolveRecallCompressorProfile picks the first available candidate by slug; falls back optimistically to the first candidate when the catalogue is missing. * invalidateRecallCompressorProfileCache() rms the persisted file. * detectRecallCompressorProfile is now cache-first and never spawns. - auto-recall.mjs: runCodexCompressor invalidates the cache on spawn error, timeout, non-zero exit, and read failure (best-effort, no error surface to user). - recall-compressor-profile.test.mjs: 11 unit tests covering catalogue read, candidate selection (with/without configured first), missing catalogue fallback, configured_off path, invalidate, cache-first detect, and re-resolve after invalidate. Notes: - buildCodexExecArgs is still exported so auto-recall can spawn the actual compress run; the change only removes the *probe* spawn, not the compress spawn. - recallCompressDetectTtlMs and recallCompressDetectTimeoutMs are preserved in config for back-compat; the timeout no longer matters but the TTL still bounds how stale a cached profile may be. * [codex] omit X-OpenViking-Actor-Peer env_http_headers when no peer configured syncMcpConfig used to unconditionally write all three OV header→env mappings. The wrapper strips empty OPENVIKING_PEER_ID before exec'ing codex, so an unset env var would silently flip the header to "" — the OV side then has to disambiguate that from "no peer scope". Match the bearer_token_env_var pattern: present only when there's something to send. Also drops a stale X-OpenViking-Actor-Peer entry when the peer is unset (e.g. after switching ovcli configs). - Existing test 4 became two cases: with-peer keeps the mapping, without-peer drops it (symmetric to bearer). - New test asserts an in-place drop when the cached .mcp.json had a stale peer mapping but the active config no longer has a peer. * [codex] runtime_failed compressor marker stops same-session retry storms Previous fix invalidated the profile cache on compress failure. Within a single codex session that still bled `recallCompressTimeoutMs` of wall time per UserPromptSubmit because the next hook reread cache (miss), fell back to fallbackRecallCompressorProfile, and tried the same model. Replace plain invalidate with a runtime_failed sentinel cached in the profile slot. UserPromptSubmit's compressMemoryContext already short- circuits on `profile.enabled === false`, so the marker stops further spawns for the rest of the codex process. The next SessionStart cache- first detect treats `source === 'runtime_failed'` as cache miss and re-resolves against the current models_cache.json, so a transient failure self-recovers across codex restarts without operator action. detect_on_startup=false respects the marker (no auto-recover, matches the "manual control" intent of that flag). - recall-compressor-profile.mjs: * markRecallCompressorRuntimeFailed(cfg, {failedModel}) writes the disabled sentinel. * detectRecallCompressorProfile branches on cached.source === 'runtime_failed': cache hit otherwise, recover-via-resolve when startup-detect on, respect marker when off. - auto-recall.mjs::runCodexCompressor: swap invalidate-on-error with markRecallCompressorRuntimeFailed(cfg, {failedModel: profile.model}). - recall-compressor-profile.test.mjs: 4 new tests covering marker write, cross-restart recovery picking a different slug, and the detect_on_startup=false honor path. 20/20 pass. invalidateRecallCompressorProfileCache is kept as a public API for explicit operator use (e.g. a future `ov codex reset-compressor` command), but is no longer called from the runtime path.	2026-06-16 02:05:18 +08:00
Qin Haojie	ff258768c2	feat(memory): 引入 User/Peer 记忆隔离模型 (#2236) ... * feat(memory): introduce user and peer memory isolation Unify agent-scoped memory behavior into user-owned memory spaces, add peer_id compatibility for session and retrieval paths, and wire memory_policy through session commit flows. * feat(memory): align session identity around peer IDs * feat(search): pass peer id through retrieval * refactor(memory): remove agent identity from integrations * fix(memory): isolate peer identity from self extraction * fix(tau2): provision benchmark user configs * fix(auth): allow admin keys to access data APIs * fix(openclaw): enable peer memory policy for peer roles * fix(openclaw): resolve sender for peer recall * refactor(session): simplify memory extraction routing * refactor(ov-cli): reduce formatting-only diff * refactor(message): remove unused message helpers * refactor(retrieval): simplify peer target resolution * refactor(namespace): remove deprecated agent namespace policy * fix(agent): propagate peer id through integrations * fix(auth): align integration clients with api-key mode	2026-06-05 10:55:48 +08:00
Zayn Jarvis	a7d27920cd	fix(plugin/codex): simplify commit hook messages (#2036)	2026-05-14 13:04:01 +08:00
3 people	e92180a7e1	feat(plugin/codex): add lifecycle hooks (recall, capture, pre-compact) to codex-memory-plugin (#1957) ... * feat(plugin/codex): add lifecycle hooks (recall, capture, pre-compact) Brings the codex-memory-plugin to feature parity with the claude-code-memory-plugin by wiring the four Codex lifecycle hooks via `hooks.json`: - SessionStart -> bootstrap-runtime.mjs (npm ci into ${CODEX_PLUGIN_DATA}/runtime) - UserPromptSubmit -> auto-recall.mjs (search OV, inject via hookSpecificOutput.additionalContext) - Stop -> auto-capture.mjs (incremental transcript capture + last_assistant_message commit) - PreCompact -> pre-compact-capture.mjs (full transcript -> single OV session -> commit) Differences from the Claude Code plugin baked into the scripts: - Codex output schema does not allow `decision: "approve"`; no-op is `{}` - Stop/PreCompact only support `systemMessage`, not `additionalContext` - Plugin envs are CODEX_PLUGIN_ROOT / CODEX_PLUGIN_DATA - Config section is `codex` (was `claude_code`); config file defaults to `~/.openviking/ovcli.conf`, falling back to legacy `~/.openviking/ov.conf` Other changes: - src/memory-server.ts now reads ovcli.conf-style configs (top-level `url`, `api_key`, `account`, `user`, `agent_id`) so the plugin works against hosted OpenViking deployments out of the box. Env-var-only operation (OPENVIKING_URL set, no config file) is also supported. - .mcp.json points at scripts/start-memory-server.mjs, which boots the same runtime the hooks use, so the MCP path benefits from npm-ci bootstrap. - README rewritten with architecture diagram, validation SOP, configuration reference, and a Codex-vs-Claude-Code differences table. Validated end-to-end against an OpenViking deployment: - Auto-recall returns ranked memories with full content and emits hookSpecificOutput.additionalContext. - Auto-capture (last_assistant_message path) creates a session, commits, and the OV pipeline extracts events + preferences within ~60s. - Pre-compact-capture posts a full 4-turn transcript to one OV session, commits with archived=true, and produces structured leaf memories (preferences, events, entities) under viking://user/<user>/memories/. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(plugin/codex): drop SessionStart, split Stop=add_message vs PreCompact=commit Codex's `Stop` hook fires per turn, not at session end, so committing per-Stop over-fragments memory extraction. And codex re-fires `SessionStart` on short reconnects, so registering an `npm ci` bootstrap there reinstalls the runtime unnecessarily. This change keeps one long-lived OpenViking session per codex `session_id` across all `Stop` invocations, and only triggers the OV memory extractor on `PreCompact` (or via an idle-sweep best-effort commit when codex exits without compacting). - hooks.json: drop SessionStart entry; keep UserPromptSubmit/Stop/PreCompact - scripts/session-state.mjs (new): per-codex-session state under ~/.openviking/codex-plugin-state/, tracks ovSessionId + capturedTurnCount - scripts/auto-capture.mjs (Stop): incremental add_message only, idle-sweep at the tail to commit stale codex sessions (default IDLE_TTL=30 min, override with OPENVIKING_CODEX_IDLE_TTL_MS) - scripts/pre-compact-capture.mjs (PreCompact): catch-up append + commit the long-lived OV session, then null out ovSessionId so the next Stop opens a fresh OV session for the post-compact half - MCP runtime install stays lazy in start-memory-server.mjs (already there); no SessionStart hook means short reconnects don't re-trigger npm ci - VERIFICATION.md: end-to-end SOP against a live OV server (~3 min) - bump plugin to 0.3.0 Verified end-to-end against ov.zaynjarvis.com: Stop adds turns idempotently and incrementally; PreCompact commits to history/archive_001/ with extractor producing memories under viking://user/<user>/memories/profile.md after ~30 s; post-compact Stop opens a fresh OV session; idle-sweep commits stale state files. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(plugin/codex): replace idle-sweep with SessionStart(source=clear) commit Per Zayn's followup ("非必要不要加 idle commit"): drop the idle-sweep added in the previous commit and use codex's actual context-disappearing signal — SessionStart with source=clear — to commit orphaned sessions. Codex hook signal map: - /compact → PreCompact ✅ commit (already) - /clear → SessionStart(source=clear) for the NEW session_id; the prior transcript is orphaned. Now committed. - /new → SessionStart(source=startup); ambiguous with fresh codex startup, so we don't act on it. - /resume / short reconnect → SessionStart(source=resume|startup); no-op to avoid corrupting still-active sessions. - SIGTERM/Ctrl+C/exit → no hook fires. Documented as a known gap; users should /compact before /exit if they want commit. Changes: - new scripts/session-start-commit.mjs: gates internally on source=clear, iterates listStates(), and commits any state file whose codexSessionId != the new SessionStart session_id, then clears that state file - hooks/hooks.json: re-register SessionStart pointing at the new script (timeout 30s) - scripts/auto-capture.mjs: remove sweepIdleSessions() and IDLE_TTL_MS env handling; Stop is now strictly add_message - README/VERIFICATION.md: update arch diagram, replace idle-sweep step with SessionStart(source=clear) verify (positive + negative paths), add "Known gap: SIGTERM/exit are silent" section - bump to 0.3.1 Verified end-to-end against ov.zaynjarvis.com: Stop add+idempotent ✓ SessionStart source=startup → {} ✓ SessionStart source=resume → {} ✓ SessionStart source=clear → committed prior OV session, history/archive_001/ appeared, profile.md gained "Favorite snack: dark chocolate" within 30 s. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore(plugin/codex): SessionStart matcher = "clear" (native dispatcher gate) Codex's hooks dispatcher matches the SessionStart hook's `matcher` field against the SessionStart `source` value. Setting matcher to "clear" means codex won't even spawn our script on `source=startup` or `source=resume` (short reconnects); we previously gated this in-script. The internal source check in session-start-commit.mjs is kept as defense-in-depth. Source: codex-rs/hooks/src/events/session_start.rs `select_handlers(..., matcher_input: Some(request.source.as_str()))` and codex-rs/hooks/src/events/common.rs `is_exact_matcher` — "clear" is all-alphanumeric so it's matched as exact equality, not regex. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(plugin/codex): active-window heuristic + idle-TTL sweep at SessionStart (v0.4.0) Source of truth: examples/codex-memory-plugin/DESIGN.md (added in this commit). Behavioral changes: - SessionStart matcher widens from `clear` to `clear|startup`. Both sources run the same active-window heuristic; `resume` is a hard no-op (still fires on short reconnects). - Heuristic (DESIGN.md §3): count state files (excluding new session_id) within ACTIVE_WINDOW_MS (default 2 min). 0 → noop, 1 → commit it (just-ended session), ≥2 → skip and rely on idle TTL. Tunable via OPENVIKING_CODEX_ACTIVE_WINDOW_MS. - Idle-TTL sweep returns at the tail of session-start-commit.mjs only (not every Stop). Default IDLE_TTL_MS = 30 min via OPENVIKING_CODEX_IDLE_TTL_MS. Catches SIGTERM/Ctrl+C/`/exit` orphans and the ≥2-active skip path. - Stop hook deliberately does NOT sweep — state-write-on-every-turn already gives us the freshness signal. Marker comment added. - Stop hook adds post-compact transcript-shrink defense: if allTurns.length < state.capturedTurnCount, reset capturedTurnCount = 0. - Commit-on-failure preserves state everywhere (PreCompact, heuristic, idle sweep). A non-2xx /commit no longer clears ovSessionId; the next sweep retries. - session-state.mjs saveState now uses atomic write (tmpfile + rename) for crash safety. listStates ignores the brief `<id>.json.tmp` window. Bump: package.json + .codex-plugin/plugin.json → 0.4.0. Docs: README "How It Works" gained a DESIGN.md pointer and rewrites the SessionStart section to reflect heuristic + idle TTL. VERIFICATION.md step 6 now exercises all four heuristic branches (0/1/≥2 active, idle TTL, resume). Phase-2 resume context inject documented in DESIGN.md but explicitly out of scope here. Verified locally with synthetic stdin tests against a fake OV server: 1-active commit, ≥2-active skip, idle TTL sweep, resume noop, unreachable-server keeps state. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(plugin/codex): align config loading with claude-code plugin Addresses three review points on PR #1957: 1. Honor OPENVIKING_CLI_CONFIG_FILE for the ovcli.conf override path (matches the convention used by `ov` CLI and claude-code-memory-plugin). OPENVIKING_CONFIG_FILE stays as the ov.conf override; for backward compat it still works when pointed at an ovcli-shaped file. 2. Strict env-first priority for every connection / identity field (baseUrl, apiKey, account, user, agentId). Env vars now win over ovcli.conf, which wins over ov.conf's codex.* block / server.*, which wins over built-in defaults. 3. Unify hook and MCP-server config loading: src/memory-server.ts now imports loadConfig from scripts/config.mjs (relative path stays valid post-compile because servers/ and scripts/ are siblings), eliminating the divergent account/user/agentId fallback chains the PR-Agent reviewer flagged. Auth header: emit Authorization: Bearer (primary, required by OpenViking Cloud) plus the legacy X-API-Key during the transition window. All six fetch sites updated (4 hook scripts + memory-server.ts + compiled servers/memory-server.js). README: document the new resolution chain, OPENVIKING_CLI_CONFIG_FILE, OPENVIKING_BEARER_TOKEN alias, and the Authorization: Bearer migration. * docs(plugin/codex): put installation first * fix(plugin/codex): harden runtime and capture paths * docs(plugin/codex): align local marketplace name * docs(plugin/codex): add one-line installer * fix(plugin/codex): support branch installer testing * fix(plugin/codex): keep installer env surface stable --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> Co-authored-by: zhengxiao.wu <zhengxiao.wu@bytedance.com>	2026-05-13 19:25:30 +08:00