v2.11.7

[2.11.7] — 2026-09-18

Added

• The org bus now carries a tool_result event when a tool call completes, so "did that command work?" is a field rather than an inference from the agent's own narration — a Bash running a test suite previously looked identical on the bus whether it passed, failed, or the binary was missing. Correlated to its invocation by the SDK's per-call id (so two concurrent Bash calls from one role stay distinct), carrying ok, duration and output capped at 4,000 characters with redactSecrets applied before the cut and truncation signalled structurally. Typed as ToolResultEventData in types.ts rather than an ad-hoc literal; runners that cannot observe tool completion simply never emit it (#289: a4bcf86).
• Decision traces carry a structured kind (fence-block, gate-pending, policy-deny, approval-pending, approval-resolved, cross-org-handoff, …). A prompt-injection fence block and a routine wait for human approval previously emitted identical structured fields, distinguishable only by matching English prose. The field is required in recordDecision's signature, so the compiler guarantees every emitter populates it — which covered four emitters, not the two originally reported (#290: 071a646).

Fixed

• Org run memory was silently dropped when the memory backend could not load: bridgeStoreEntry returns null, storeRunMemory ignored it, and its caller's catch never fired because nothing threw. Runs completed normally, runtime.json and history were written, the bus looked perfect — and every org_recall came back empty, with the only trace printed under MONOMIND_DEBUG=1. A failed store now surfaces three ways that outlive the terminal: an org-memory-store-failed audit event, an unconditional warning, and a memoryError field in runtime.json that org status prints (human and --format json). It deliberately does not throw — by then the run has succeeded and its history is on disk, so throwing would fail a run that worked and blame history for it. The same swallowed-null pattern was fixed in four more callers, two of which actively misreported: hooks post-command returned recorded: true for a write that never happened and skipped its JSON fallback (losing the record twice), and a consolidation worker counted patterns it never wrote (#293: bc75ea5).
• policy.git denials now name the boundary and the allowed alternative, not just the rejected attempt. "path escapes org workdir" never said what the workdir was, so a role could only guess another path — in one rehearsal a reviewer's single Read was denied, it never learned the root it ran under, and it reviewed from submission messages without reading a line of code. Five denial messages fixed: workdir escape, write-scope, path-less Grep/Glob, tool allowlist and research-domain allowlist (#291: 3911a5e).
• monomind org branch <org> <run> <label> read as though <label> named the new run; the id is generated, and the label was not merely a note — it was discarded entirely, never reaching .branch-source. The label is now recorded, --format json prints the generated id ({"v":1,"org":…,"run":…,"from":…,"label":…}) so a script can replay without parsing prose, and the help no longer shows a label in the id position. The label is deliberately NOT used as the run id: run ids are joined into filesystem paths and the codebase already guards that shape against traversal (#292: f4c08a3).
• biome linted nothing inside .claude/worktrees, which is where this repo's own workflow puts worktrees: npx biome check reported "Checked 0 files" and naming a file said the path was ignored. A "lint is clean" reading was really "biome refused to look" — CI was unaffected, but anyone working the recommended way was misled. The ignore pattern now excludes the .claude assets without excluding a worktree checkout's own source, with a repo test asserting both directions (#294: 71b3e17).

Changed

• scripts/sync-claude-trees.mjs (pnpm run sync:claude-trees, with --check in verify and CI) keeps the five .claude asset trees canonical after init marks files in this repo. It normalises rather than mirrors: the shipped packages/@monomind/cli/.claude is init's asset source, so copying the marked root copy into it would ship per-project markers to every npm user and make the next init nest a second block inside the first. It never creates and never deletes, which is how the shipped superset is protected structurally. The dead sync-claude-assets.sh (a hard exit 1 since July, still referenced by three checklists) is deleted, and the parity test, skill lint and doc/publishing.md now name the real command (42cb52e).
• @monoes/hooks 1.0.6 → 1.0.7 (the consolidation-worker fix above).

npm: monomind@2.11.7 · @monoes/monomindcli@2.11.7 · @monoes/hooks@1.0.7