jonah-fleet 1.36.0 → 1.37.1

package/CHANGELOG.md

@@ -5,6 +5,26 @@ All notable changes to `jonah-fleet` will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.37.1](https://github.com/Jonah-Projects/jonah-fleet/compare/v1.37.0...v1.37.1) (2026-10-12)
+
+
+### Bug Fixes
+
+* **prompts:** prevent autowork repetition polling loops on in-flight peer-review PRs ([#713](https://github.com/Jonah-Projects/jonah-fleet/issues/713)) ([ceb76f8](https://github.com/Jonah-Projects/jonah-fleet/commit/ceb76f84c78b16d9075b1b4e453c6ae8696c1d88))
+
+## [Unreleased]
+
+### Fixed
+
+- **prompts:** prevent autowork repetition polling loops on in-flight peer-review PRs, define the Active Peer Review & Non-Draft Awaiting Review Invariant in `autowork.md`, and protect `Starting review (round N)` claim windows in `ORCHESTRATION.md`
+
+## [1.37.0](https://github.com/Jonah-Projects/jonah-fleet/compare/v1.36.0...v1.37.0) (2026-10-11)
+
+
+### Features
+
+* **prompts,memory:** partition LESSONS.md ingestion by routine role ([#700](https://github.com/Jonah-Projects/jonah-fleet/issues/700)) ([21fe1b2](https://github.com/Jonah-Projects/jonah-fleet/commit/21fe1b21669ee0d7dc470c601732fb395efa0126))
+
 ## [1.36.0](https://github.com/Jonah-Projects/jonah-fleet/compare/v1.35.2...v1.36.0) (2026-10-11)
 
 
@@ -19,16 +39,6 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 * **terminal-card:** refine backlog triage hints for radar digests and needs-attention issues ([#692](https://github.com/Jonah-Projects/jonah-fleet/issues/692)) ([935ef75](https://github.com/Jonah-Projects/jonah-fleet/commit/935ef75a5e4cbd607bc75f0767901f6f6d7927a8))
 
-## [Unreleased] - 2026-10-11
-
-### Added
-
-- **runner:** add head-tail streaming truncation circuit breaker (`DEFAULT_MAX_LINES = 500`, `DEFAULT_MAX_BYTES = 64KB`) for command execution and large tool outputs to prevent context overflow and worker stalls
-
-### Fixed
-
-- **terminal-card:** refine backlog triage hints to provide targeted unblock advice for radar digests and `needs-attention` issues instead of generic `needs-human` prompts
-
 ## [1.35.1](https://github.com/Jonah-Projects/jonah-fleet/compare/v1.35.0...v1.35.1) (2026-10-10)
 
 

package/dist/lib/memory.d.ts

@@ -0,0 +1,61 @@
+/**
+ * Operational Memory (LESSONS.md) Partitioning and Filtering
+ *
+ * Implements role-filtered memory recall inspired by huggingface/funes v1.6.0 (PR #214),
+ * allowing autonomous routines (autowork, peer-review, optimizer, etc.) to ingest
+ * only role-relevant and general lessons rather than the unbounded file.
+ */
+export declare const KNOWN_ROUTINE_ROLES: Set<string>;
+export interface LessonEntry {
+    raw: string;
+    role: string;
+    tag: string;
+    title: string;
+    symptom?: string;
+    rootCause?: string;
+    rule?: string;
+}
+export interface LessonEntryInput {
+    role?: string;
+    tag?: string;
+    title: string;
+    symptom: string;
+    rootCause: string;
+    rule: string;
+}
+/**
+ * Parses raw LESSONS.md content into header preamble and array of structured entries.
+ */
+export declare function parseLessonsFile(content: string): {
+    preamble: string;
+    entries: LessonEntry[];
+};
+/**
+ * Parses structured lesson entries from LESSONS.md body or full file content.
+ */
+export declare function parseLessonEntries(content: string): LessonEntry[];
+/**
+ * Checks whether a lesson entry matches a given routine role.
+ * Includes:
+ * - Direct matches (e.g. entry tagged 'autowork' when role is 'autowork')
+ * - Composite matches (e.g. entry tagged 'autowork:git')
+ * - 'general' entries
+ * - General/legacy subsystem entries not claimed by another known routine role
+ */
+export declare function isLessonRelevantForRole(entry: LessonEntry, role: string): boolean;
+/**
+ * Filters LESSONS.md content to only entries relevant for the active routine role.
+ */
+export declare function filterLessonsByRole(content: string, role: string): string;
+/**
+ * Formats a structured 3-line lesson entry under the active routine tag.
+ */
+export declare function formatLessonEntry(input: LessonEntryInput): string;
+/**
+ * Appends a new structured lesson entry to LESSONS.md, enforcing the 25-entry hard cap.
+ * If adding the entry exceeds maxEntries, the oldest entries are evicted.
+ */
+export declare function appendLessonEntry(content: string, input: LessonEntryInput, options?: {
+    maxEntries?: number;
+}): string;
+//# sourceMappingURL=memory.d.ts.map

package/dist/lib/memory.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"memory.d.ts","sourceRoot":"","sources":["../../src/lib/memory.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,eAAO,MAAM,mBAAmB,aAS9B,CAAC;AAEH,MAAM,WAAW,WAAW;IAC1B,GAAG,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,MAAM,CAAC;IACb,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,gBAAgB;IAC/B,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;IAClB,IAAI,EAAE,MAAM,CAAC;CACd;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,WAAW,EAAE,CAAA;CAAE,CAc9F;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,WAAW,EAAE,CA4DjE;AAED;;;;;;;GAOG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,WAAW,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAuBjF;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,MAAM,CAczE;AAED;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,gBAAgB,GAAG,MAAM,CAGjE;AAED;;;GAGG;AACH,wBAAgB,iBAAiB,CAC/B,OAAO,EAAE,MAAM,EACf,KAAK,EAAE,gBAAgB,EACvB,OAAO,GAAE;IAAE,UAAU,CAAC,EAAE,MAAM,CAAA;CAAO,GACpC,MAAM,CAoBR"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jonah-fleet",
-  "version": "1.36.0",
+  "version": "1.37.1",
   "description": "Standalone autonomous agent fleet with Symphony orchestration, claim protocols, and continuous improvement loops",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",

package/templates/docs/AGENTS.template.md

@@ -98,3 +98,10 @@ To trigger the Autowork routine manually on any specific issue, use the GitHub A
 - Auth boundaries required for all sensitive mutations.
 - Never commit credentials, tokens, or secret keys to version control.
 - Validate request inputs and sanitize outputs.
+
+## Operational Memory (LESSONS.md)
+
+- **Role-Partitioned Memory**: Standardize lesson entries in `LESSONS.md` using routine role tags: `[autowork]`, `[peer-review]`, `[optimizer]`, or `[general]` (e.g. `### [autowork] Title`, `### [peer-review] Title`, `### [general] Title`). Subsystem composite tags are also supported (e.g. `### [autowork:subsystem] Title`).
+- **Targeted Routine Ingestion**: Autonomous routines ingest general lessons plus routine-specific lessons, omitting unrelated entries to prevent unbounded token expansion while maintaining high-signal operational memory.
+- **Post-Run Recording**: When capturing new non-obvious gotchas or quirks on PR branches, always tag the entry with the active routine's role (or `[general]`) and adhere strictly to the 25-entry hard cap.
+

package/templates/docs/LESSONS.template.md

@@ -1,7 +1,9 @@
 # Operational Lessons & Repository Heuristics
 <!-- Invariants: Max 25 entries. Hard cap. Older entries graduate or demote to LESSONS_ARCHIVE.md -->
+<!-- Tagging Syntax: Standardize entries with routine role tags [autowork], [peer-review], [optimizer], or [general] (e.g. ### [autowork] Title, ### [peer-review] Title, ### [general] Title). Subsystem suffixes are supported: ### [autowork:subsystem] Title. Each routine ingests only its own role-tagged entries plus [general] and legacy entries. -->
 
 ### [subsystem] Title
+<!-- Example role-tagged entry: ### [general] Title or ### [autowork] Title -->
 - **Symptom:** <Symptom description or error message>
 - **Root Cause:** <Brief explanation of the underlying cause>
 - **Rule:** <Actionable rule or constraint to follow>

package/templates/prompts/ORCHESTRATION.md

@@ -28,7 +28,7 @@ Autowork triggers pass the target issue via environment variables (`TARGET_ISSUE
 
 Invariants deliberately upheld from this spec:
 
-- **Single-flight per issue and PR convergence** — at most one run works an issue or pull request at a time, enforced by the autowork claim protocol (assign → read-back → earliest-timestamp tiebreak). For **umbrella** issues, single-flight is maintained at the _child-issue_ level so slices progress cleanly.
+- **Single-flight per issue and PR convergence** — at most one run works an issue or pull request at a time, enforced by the autowork claim protocol (assign → read-back → earliest-timestamp tiebreak). For **umbrella** issues, single-flight is maintained at the _child-issue_ level so slices progress cleanly. Pull requests actively claimed by peer review (`Starting review (round N)` within 2 hours) or awaiting review are strictly protected from concurrent autowork convergence and repetitive polling loops.
 - **Recover dead-run claims** — a crashed run's orphaned claim is released back to the pool rather than starving the issue or PR, both opportunistically during candidate selection and periodically via issues housekeeping.
 - **Reader/writer separation** — the routine that authors a PR never merges it; the Peer Review routine is the sole merge authority for pull requests.
 - **Warm-Context Review Synchronization** — Autowork maintains an active warm session during implementation, polling for Peer Review's verdict. When Peer Review bounces a PR to draft with findings, Autowork immediately detects the draft state in-session, applies fixes directly to its warm working tree, and re-marks the PR ready—re-firing Peer Review for Round N+1 without cold-start overhead.
@@ -54,7 +54,7 @@ Single source of truth for both autowork candidate reclamation and housekeeping
 
 Single source of truth for Autowork Phase 1 pull request convergence. An assigned pull request or draft PR with unaddressed review comments is a _stale claim_ (safe to reclaim and reassign by another runner) only when **all** of these hold:
 
-1. **It carries an autowork claim comment**: The PR thread contains `🔒 Addressing review findings by autowork run …` or `🔒 Addressing review findings by local autowork session …`.
+1. **It carries an autowork claim comment or peer-review marker**: The PR thread contains `🔒 Addressing review findings by autowork run …`, `🔒 Addressing review findings by local autowork session …`, or `Starting review (round N) …`.
 2. **The claim is old**: The most recent claim comment's `created_at` is **more than 2 hours** ago. (2 hours instead of 6 hours because PR review convergence is a rapid turnaround loop).
 3. **No active commits or review activity**: No new commit has been pushed to the PR branch within the last 2 hours.
 

package/templates/prompts/autowork.md

@@ -61,6 +61,7 @@ If any criterion cannot be met, stop immediately and log FAILURE with the reason
 - Do not fall into the **Telemetry Rabbit Hole**: do not spend cycles instrumenting elaborate fallback telemetry or defensive error handling for features that suffer from lack of user intent rather than software bugs.
 - Do not guess or invent arbitrary specifications for ambiguous issues — post the 4-part escalation card ('Why I believe this'), label `needs-info`, and release the claim instead of blindly writing code.
 - Do not call `schedule` or yield the turn with plain text while waiting for background verification tasks — stay in the tool loop until the Definition of Done is met.
+- Do not busy-poll external routine issues, PRs under review, or background tasks in a repetitive tool loop; never poll in a loop with identical commands (e.g. repeated `gh issue view <peer_review_run>` or `gh run view`) without making progress.
 - Do not write implementation code before proving that the reproduction test or verifier fails (the Red Gate).
 - Do not pause or prompt rhetorical questions ("Shall I continue?", "Should I start implementing?")—proceed autonomously.
 - Do not use bare `git commit` or blind `git add .` without scoping specific paths (`git commit -m "..." -- <paths>`).
@@ -92,9 +93,10 @@ b. **Eligible → claim, then implement.** Call `get_me` once to learn your own
 ### Phase 1: Converge on open work (Scan mode; skipped in Targeted mode)
 
 1. List all open PRs needing attention (specifically draft PRs bounced back by peer review or open PRs with review comments: `gh pr list --state open --json number,title,isDraft,headRefName,assignees,labels,author`, excluding automated release PRs `release-please--*` / `chore(main): release*`).
+   - **Active Peer Review & Non-Draft Awaiting Review Invariant**: In Phase 1, only PRs that have been **bounced to draft** with actionable review findings, or PRs with **unaddressed review comments** (threads from peer review requiring code changes), require convergence. If a PR is ready for review (`isDraft: false`), or carries an active peer-review claim comment (`Starting review (round N)` posted within the last 2 hours), or has no unaddressed review comments / changes requested, you DO NOT OWN THIS PR. You MUST NOT attempt to address it, MUST NOT block, and MUST NEVER poll its review run issue (e.g. `gh issue view <peer_review_run>`). Skip this PR immediately and evaluate the next PR or proceed directly to Phase 2.
    1a. **PR Claim Protocol (Single-Flight PR Convergence)**:
    - For each open PR needing attention (bounced to draft with actionable findings, or open with unaddressed review comments):
-   - **Check eligibility**: re-read candidate PR (`gh pr view <PR> --json assignees,comments`). Skip PRs currently assigned to another live runner or carrying an active claim comment posted within the last 2 hours (unless stale claim per `ORCHESTRATION.md`).
+   - **Check eligibility**: re-read candidate PR (`gh pr view <PR> --json assignees,comments`). Skip PRs currently assigned to another live runner or carrying an active claim comment posted within the last 2 hours per the invariant above (unless stale claim per `ORCHESTRATION.md`).
    - Call `get_me` once to learn your own login if not already known.
    - **Claim atomically**: assign yourself to the PR (`gh pr edit <PR> --add-assignee <login>`) AND post a claim comment on the PR:
      - In **Cloud Actions**: `🔒 Addressing review findings by autowork run {run_url} {timestamp}`
@@ -220,7 +222,7 @@ _Generated by [Antigravity](${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/ru
         `- **Prior Failure Context**: <summary of past error/crash and strategy to avoid repeating it>`
     - If bug: use `/diagnosing-bugs` to establish reproduction test before fixing.
     - If large/complex: use `/domain-modeling` and `/codebase-design`.
-    - **Pre-Flight Memory Scan**: If `LESSONS.md` exists, grep matching subsystem tags (`grep -E "^### \[(subsystem)\]" LESSONS.md -A 4`) to incorporate known landmines into implementation plans before writing code.
+    - **Pre-Flight Memory Scan**: If `LESSONS.md` exists, grep matching subsystem tags (`grep -E "^### \[(subsystem)\]" LESSONS.md -A 4`) or filter entries tagged for `[autowork]` or `[general]` (e.g. `grep -E "^### \[(autowork|general)\]" LESSONS.md -A 4`), omitting unrelated routine roles to incorporate known landmines into implementation plans before writing code without token bloat.
     - **The Red Gate Protocol (Pre-Implementation Verifier Check)**: Before writing ANY production implementation code, define the verifier or write a reproduction test, and execute it:
       - **The check MUST fail** (red). If the verifier passes before you touch production code, the subsequent work proves nothing.
       - **Already Green Check**: If the verifier passes *before* code changes, stop immediately and evaluate:
@@ -268,7 +270,7 @@ _Generated by [Antigravity](${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/ru
     - **Diagnostic Reflex**: On unexpected test/build failure during TDD, grep symptom text in `LESSONS.md` before making speculative code edits.
     - Run repository tests and verification.
     - **Disk Read-Back Verification**: Verify written configs, schemas, and generated files by reading values directly back from disk files (`cat`, `jq`, test assertions), never from shell variables or memory.
-    - **Pre-PR Lessons Capture Gate**: If solving the issue required overcoming a non-obvious quirk not caught by tests/linters, append a 3-line structured entry on the active PR branch adhering to the 25-entry hard cap. Never write to `LESSONS.md` directly on `main`.
+    - **Pre-PR Lessons Capture Gate**: If solving the issue required overcoming a non-obvious quirk not caught by tests/linters, append a 3-line structured entry tagged under the active routine role (`### [autowork] <Title>` or `### [general] <Title>`) on the active PR branch adhering to the 25-entry hard cap. Never write to `LESSONS.md` directly on `main`.
     - **Milestone 2 (Verification & Tests)**: Once implementation passes tests and type checks, emit milestone card if `$ROUTINE_ISSUE_NUMBER` is set:
       ```bash
       gh issue comment "$ROUTINE_ISSUE_NUMBER" --body "### 🧪 Milestone: Verification & Tests
@@ -365,7 +367,7 @@ Cacheable across consecutive runs within the same repository.
 In this tier, Autowork ingests semi-stable repository conventions, operational memory, and configuration rules that change infrequently across consecutive runs in the same repository:
 
 1. **Repository Conventions (`AGENTS.md`)**: Read `AGENTS.md` (or `CLAUDE.md` / `GEMINI.md`) for build/test/lint commands, code formatting rules, architectural guidelines, and design system tokens.
-2. **Operational Memory (`LESSONS.md`)**: If `LESSONS.md` exists, ingest active operational gotchas (`grep -E "^### \[(subsystem)\]" LESSONS.md -A 4`) to avoid known pitfalls and regression patterns (enforcing the 25-entry hard cap).
+2. **Operational Memory (`LESSONS.md`)**: If `LESSONS.md` exists, ingest role-relevant operational gotchas (`grep -E "^### \[(subsystem)\]" LESSONS.md -A 4`) filtered for the active routine role (`[autowork]` or `[general]`), omitting entries from other routine roles to avoid known pitfalls and regression patterns while minimizing token consumption (enforcing the 25-entry hard cap).
 3. **Fleet Manifest (`agents-manifest.json`)**: Ingest configured presets, model family assignments, and timeout/iteration ceilings.
 4. **Engineering Skills (`.agents/skills/`)**: Leverage domain skills (`code-review`, `codebase-design`, `diagnosing-bugs`, `domain-modeling`, `grill-me`, `resolving-merge-conflicts`, `tdd`, `triage`, `writing-for-agents`).
 

package/templates/prompts/peer-review.md

@@ -142,7 +142,7 @@ Check if `$PR_NUMBER` is set:
 1. Run `/code-review` over the diff (or delta commits if re-review) evaluating:
    - **Standards**: Conformance to `AGENTS.md` (or `CLAUDE.md`/`GEMINI.md`), conventions, and architecture.
    - **Spec Compliance**: Verification against the linked issue's deliverables (`## Tasks`), or against the PR description's summary/changes if no tracking issue is linked.
-   - **Operational Memory & Lessons Invariant**: Inspect `LESSONS.md` diffs in PRs to verify the lesson is accurate, non-trivial, follows the 3-line structured schema, and the file adheres to the 25-entry hard cap.
+   - **Operational Memory & Lessons Invariant**: Inspect `LESSONS.md` diffs in PRs to verify the lesson is accurate, non-trivial, follows the 3-line structured schema, tags the originating routine role (e.g. `### [autowork] ...`, `### [peer-review] ...`, or `### [general] ...`), and the file adheres to the 25-entry hard cap. If reviewing code that solves a non-obvious gotcha or recurring architectural trap, ensure a structured 3-line entry tagged with the originating routine role (`### [peer-review] ...` or `### [general] ...`) is appended to `LESSONS.md`.
 2. Run Security Pass: auth gates, permission checks, injection risks, sensitive credentials.
 3. **Design System & Viewport Density Pass** (if PR modifies frontend/rendered UI):
    - **Token Purity**: Check for arbitrary CSS/Tailwind sizing overrides (e.g. `text-[...px]`, `w-[...px]`) or bespoke button styling bypassing standard design system tokens.
@@ -298,7 +298,7 @@ Cacheable across consecutive runs within the same repository.
 In this tier, Peer Review ingests semi-stable repository conventions, operational memory, and configuration rules that change infrequently across consecutive runs in the same repository:
 
 1. **Repository Conventions (`AGENTS.md`)**: Read `AGENTS.md` (or `CLAUDE.md` / `GEMINI.md`) for coding standards, architectural patterns, design system token purity, and required CI / test commands.
-2. **Operational Memory (`LESSONS.md`)**: Verify that any non-obvious quirks or gotchas captured on the active PR branch adhere to the 25-entry hard cap.
+2. **Operational Memory (`LESSONS.md`)**: If `LESSONS.md` exists, ingest role-relevant operational gotchas filtered for the active routine role (`grep -E "^### \[(peer-review|general)\]" LESSONS.md -A 4`), omitting unrelated routine roles, and verify that any non-obvious quirks or gotchas captured on the active PR branch adhere to the 25-entry hard cap.
 3. **Fleet Manifest (`agents-manifest.json`)**: Ingest configured review models, round caps, and iteration ceilings.
 4. **Engineering Skills (`.agents/skills/`)**: Leverage `code-review`, `design-critique`, and `writing-for-agents` for structured evaluation.