second-brain

Reviews a change against its git history to find the regression bug class that diff-only review misses: changes that revert or re-introduce a previously-fixed bug, or contradict the historical reason a line exists. Runs once over the changed code files using git blame/log. Dispatched by the code-review-deep skill (Pass 2c) alongside the per-unit reviewers. <example> Context: code-review-deep has decomposed a PR and is fanning out reviewers. assistant: "Dispatching code-review-history-reviewer to blame-walk the changed code for reverts and re-introduced bugs." </example>

Reviews a change for unproven RUNTIME PREMISES — load-bearing assumptions about the environment (env vars, files, process state, shared state, services, platform) that the diff depends on but does not itself establish. The bug class diff-static review misses: code internally consistent with a false belief about the runtime. Runs once over the changed code files. Dispatched by the code-review-deep skill (Pass 2d). <example> Context: code-review-deep has decomposed a change and is fanning out reviewers. assistant: "Dispatching code-review-premise-reviewer to enumerate unproven runtime premises." </example>

Scores a single code-review finding on a 0–100 confidence scale by verifying it against the actual files, and suppresses findings that match a known false-positive pattern. Dispatched (one per unique finding) by the code-review-deep skill to filter false positives before posting. <example> Context: code-review-deep collected findings from per-unit reviewers and needs to filter noise. assistant: "Scoring each finding with code-review-scorer, suppressing any that match a recorded false positive." </example>

Deep-reviews a single review unit (a group of related changed files) within a GitHub change. Reads every file in the unit, diffs each against the base, follows cross-file imports, and applies a diff-scoped bug taxonomy focused on runtime-affecting bugs — especially cross-file interaction bugs that breadth-only review misses. Dispatched in parallel (one per unit) by the code-review-deep skill. <example> Context: code-review-deep has decomposed a PR into review units and is fanning out. assistant: "Dispatching code-review-unit-reviewer per unit to hunt cross-file bugs in parallel." </example>

Background dream execution agent. Runs the transcript mining + 7-phase wiki consolidation cycle (incl. SUMMARIZE + FORGET) on a staging copy of the wiki. Dispatched by the dream skill in --background mode. Mutates only the staging directory; the FORGET phase reads the live wiki read-only to score real page ages and writes a forget-manifest — actual archiving happens only on dream_accept. <example> Context: User ran /second-brain:dream --background which created drm_20260511T143022Z. assistant: "Spawning dream-runner agent to consolidate the knowledge base in the background." </example>

Show what the second-brain safety layer did this session — every PreToolUse guard verdict, every tool-return injection flag, every wiki-write decision. Reads ~/.second-brain/audit-log.jsonl produced by hooks. Read-only.

Drop unprocessed material (a file, a URL, or pasted text) into the current project's raw inbox for later refinement into wiki notes. Usage: /second-brain:capture <path|url> | "<text>" | --list | --discard <id> | --node <slug>.

In-depth multi-pass code review of a GitHub change (local checkout). Decomposes the diff into logical review units, reviews each on the best available model (docs on the Haiku model), runs an advisory architectural pass on the highest-risk units, scores findings with an FP-aware scorer, consults the second-brain for conventions and prior reviews, and records false positives. Local output by default; --comment posts to the PR.

Adversarial validation of the second-brain plugin. Questions architectural layers from a position of maximum skepticism, drilling from general to granular. Rotates focus across runs using git-change priority, history rotation, and perspective variety so different questions are asked each time.

Run a "dream" — background consolidation of the knowledge base. Mines session transcripts for missed insights, deduplicates wiki entries, builds cross-links, and stages all changes for review before applying. Equivalent to Anthropic's managed-agents Dreams API, running locally. Supports inline (default) and background (--background) execution modes.

21 hooks across 8 events