Skip to content

project-learning.md

Latest commit

 

History

History
92 lines (70 loc) · 5.28 KB

project-learning.md

File metadata and controls

92 lines (70 loc) · 5.28 KB
Raw

Project Consolidation And Learning

Brigade keeps project consolidation and self-learning local, explicit, and reviewable.

Project Consolidation

Configure safe local project records in gitignored .brigade/projects.toml.

[[project]]
id = "project-alias"
label = "Project Alias"
category = "workflow helper"
decision = "bake-in"
reason = "Small workflow primitive belongs directly in Brigade."
docs_ready = true
license_ready = true
security_ready = true
release_ready = true
ownership_ready = true
migration_blockers = []

Supported decisions:

  • bake-in
  • integrate
  • catalog-only
  • move-candidate
  • leave-alone

Commands:

brigade projects audit
brigade projects audit --json
brigade projects readiness plan
brigade projects readiness record
brigade projects readiness list
brigade projects readiness show latest
brigade projects closeout --status deferred --reason "reviewed externally"
brigade projects closeouts
brigade projects closeout-show latest
brigade projects import-issues

brigade projects readiness plan calculates decision-specific readiness for docs, license, security, release, ownership, and migration blockers. record writes a local receipt under .brigade/projects/readiness/, while list and show inspect those receipts. Release readiness and the operator center can reference the latest receipt, but Brigade does not run any migration command.

brigade projects closeout writes reviewed project migration closeouts under .brigade/projects/closeouts/. reviewed, deferred, and archived closeouts quiet unchanged readiness issues, while changed source fingerprints resurface in doctor, brief, and import routing. superseded records that a prior closeout is no longer valid and does not quiet the active issue.

Migration plans, readiness receipts, and closeouts are manual-only. Brigade does not transfer repos, archive repos, change visibility, push, tag, publish, or mutate remotes.

Learning Loop

brigade learn aggregates local learning candidates from pending scanner imports, failed review receipts, and failed portable tool run receipts.

Commands:

brigade learn plan
brigade learn doctor
brigade learn import-issues
brigade learn import-learnings
brigade learn import-learnings --file .learnings/ERRORS.md --dry-run
brigade learn skill-candidates --source security-scan
brigade learn propose-skill <candidate-id> --dry-run
brigade learn propose-skill <candidate-id>
brigade learn closeout <candidate-id> --status accepted-risk --reason "reviewed"
brigade learn closeouts
brigade learn closeout-show latest
brigade learn replay export scenario-id --before-summary "before" --after-summary "after"
brigade learn replay list
brigade learn replay show latest
brigade learn replay compare latest

brigade learn import-learnings reads structured .learnings/ markdown logs that some operator workflows already keep on disk and proposes one local work-import per logged entry. Each entry is a level-two heading carrying a typed id such as ERR-20260311-001, LRN-20260311-001, or FEAT-20260311-001, with optional **Priority**, **Status**, and **Area** fields and free-form prose. ERR entries become incident imports, LRN entries become finding imports, and FEAT entries become feature task imports. By default the importer scans .learnings/ERRORS.md, .learnings/LEARNINGS.md, and .learnings/FEATURE_REQUESTS.md; pass --file to override. Entry text, fields, and prose are redacted and prompt-injection-guarded before they become imports, and the importer never edits the log files or canonical memory. Imported entries flow through learn plan, learn doctor, and learn skill-candidates, so repeated entries that share a logged area surface as one reviewed promote-candidate.

Every candidate should end in one reviewed path:

  • task
  • Memory Handoff draft
  • suppression or accepted risk
  • archive or dismissal

Learning closeouts support accepted-risk, dismissed, archived, and deferred. Unchanged closed-out candidates stop making the learning queue noisy, while changed source fingerprints reappear in learn plan, learn doctor, and learn import-issues.

brigade learn skill-candidates groups repeated learning evidence into reviewed skill candidates. Use --source security-scan to focus on repeated security findings such as exposed credentials in session transcripts. Candidates explain why evidence was grouped, carry review risk, preserve redacted response options, and stay manual-only. brigade learn propose-skill <candidate-id> --dry-run previews generated source and planned inbox writes. Without --dry-run, it writes source under .brigade/learn/skill-workshop/ and submits it to .brigade/skills/inbox/ using the normal skill proposal flow. The proposal is never installed automatically; review it with brigade skills inbox show, diff, accept, or reject.

Learning replay receipts store safe before/after summaries under .brigade/learn/replays/. brigade learn replay compare writes a compare receipt under .brigade/learn/replay-compares/, reports improved, unchanged, regressed, or unknown outcomes, and surfaces regressions in release evidence and operator-center reviews.

Learning receipts store safe summaries only. Brigade does not edit canonical memory, source files, tool configs, or policies automatically.