Repository navigation
chore(main): release 0.1.4 - #2
Merged
boadij merged 1 commit intoSep 7, 2026
Merged
Conversation
boadij
deleted the
release-please--branches--main--components--pi-herdsman
branch
September 7, 2026 12:45
This was referenced Sep 21, 2026
This was referenced Sep 29, 2026
boadij
added a commit
that referenced
this pull request
Oct 1, 2026
## Implementation report - **Base:** `main` at `156b1c661a2e147d6bb2ef415abe44dd6b11af2f` - **Head:** `ba06fdfe82f2dfce80aacec7d6cd09589db53f57` (`fix: align project coordination handoff behavior`) - **PR:** #201 ### Project coordination - Added automatic nonterminal Manager handoff when an assigned project Lead herd run settles. - The handoff reuses ordinary durable `ProjectMessage` transport and persists independently of current Manager availability. - The latest meaningful Lead assistant response from the herd run is included when available: `Herd run settled.\n\n<summary>` - Automatic handoff does **not** resolve the project assignment. Only Manager `staff_complete` or `staff_discard` is terminal. - Project publication and terminal resolution share the canonical fail-fast project-assignment lock so a late handoff cannot recreate messages after resolution. ### Supervisor awareness - Added transient Lead supervisor-state context derived from current Pi/Herdr authority rather than persisted presence. - Assigned project Leads distinguish: - verified live Manager → `available` - verified absent Manager → `unavailable` - inconclusive Manager verification → `unknown` - Ordinary Leads retain fail-closed behavior when no supervisor can be verified. - Supervisor-state messages are hidden, semantic, and appended only when the observed state changes; Manager identity is not exposed to the model. - Tool definitions remain stable. ### Coordination guidance - Updated `SKILL.md` and coordination docs so normal project progress stays local and Herdsman performs the routine settled-run handoff automatically. - `supervisor_message` is reserved for coordination that changes what the supervisor needs to decide, act on, review, or know about, including blockers, warnings, scope changes, risks, and explicit evidence. - Assigned project messages remain retained for the Manager role across Manager absence or turnover. - ADR 0012 now states that settled project herd runs may emit nonterminal Manager handoffs and that these do not resolve the assignment. ### Validation - Focused suites passed: - controller-lifecycle: **93** - controller-api: **93** - commands: **75** - extension-contract: **41** - supervision: **54** - `npm run check`: **779 passed, 1 skipped, 0 failed** - `npm run build`: passed - `npm run package:audit`: passed (**54 files**) - `prettier . --write`: passed - `git diff --check`: passed - `npm run smoke -- manager-recovery`: passed on `ba06fdfe82f2dfce80aacec7d6cd09589db53f57` - Recorded environment: Node **26.8.1**, Pi **0.99.2**, Herdr client/server **0.9.3** - Smoke cleanup passed and the worktree remained clean. ### External review - **External Review #2:** no findings at `ba06fdfe82f2dfce80aacec7d6cd09589db53f57`. - The four findings from External Review #1 were resolved: - assigned Manager verification preserves `unknown` instead of inventing `unavailable` - Lead guidance no longer requires an available supervisor for project messaging - project publication and resolution use the canonical fail-fast assignment lock - duplicate controller-api fixture field removed - GitHub Actions Validate run **#575** passed on this head across Ubuntu, macOS, Windows, package, container, scope, and aggregate CI. ### Reload behavior tracked separately A later live test showed that the same automatic project handoff works with fresh Pi Manager/Lead processes but can fail after `/reload` of existing sessions until Pi is fully restarted. That reload-specific behavior is tracked separately in **#205** and is not part of this PR's normal fresh-session handoff implementation. ### Diff - **11 files changed** - **990 additions** - **162 deletions** --------- Co-authored-by: Jeffrey Boadi <25706638+boadij@users.noreply.github.com>
This was referenced Oct 1, 2026
boadij
added a commit
that referenced
this pull request
Oct 4, 2026
## Summary - Unify management selection menus around Pi's native SelectList, SelectItem description metadata, Text contextual help, and getSelectListTheme(). - Group low-frequency configuration under Settings; flatten Definitions and align Running, layout, message-limit, and Chief interactions. - Preserve runtime, configuration, definition, identity, and lifecycle semantics. SettingsList was deliberately unnecessary: existing async command workflows remain better expressed through the shared native SelectList selector. ## Validation (current integrated PR head) - `npm test -- extension/commands.test.ts`: 90 passed. - `npm test -- extension/controller-api.test.ts`: 95 passed. - `npm run build`: passed. - `npm run check`: 835 passed, 1 skipped, 0 failed. - `npm run smoke -- manager-recovery`: passed. - `git diff --check`: passed. ## External Review #1 resolution / validation - Applied Pi's native compact SelectList layout, clarified that explicit Model/Thinking values apply to future Agent generations while unset values inherit or restore as appropriate, changed the shared footer to “Esc close,” and restored the schema Fields table formatting while retaining the managed-Lead sentence correction. - Focused checks: `npm test -- extension/commands.test.ts` 88/88; `npm test -- extension/controller-api.test.ts` 95/95; `npm run build` passed; `npm run check` 833 passed, 1 skipped, 0 failed; `git diff --check` passed. - Earlier package audit: `npm run package:audit` passed (59 files). Live smoke not run. - No Prettier rerun for this follow-up: a formatter attempt reflowed the schema table and CHANGELOG; those incidental changes were restored surgically. No permanent formatter configuration change was made. ## External Review #2 resolution / validation - Limited the duplicate Running-label ambiguity guard to RPC, where string selection cannot distinguish rows; TUI uses stable row values and retains exact fresh identity revalidation before focus. - Added managed-Lead-specific Model/Thinking help, one-row vertical spacing around contextual help, and updated Getting Started to reflect the Settings hierarchy. - Validation: `extension/commands.test.ts` 89/89; `extension/controller-api.test.ts` 95/95; `npm run build` passed; `npm run check` 834 passed, 1 skipped, 0 failed; `git diff --check` passed. - Prettier ran once with a temporary repository-root ignore excluding only the schema table and CHANGELOG; no persistent config or broad churn. Live smoke: NOT RUN because no compatible TTY/Herdr session was available. ## External Review #3 resolution / validation - Removed the shared 32-column cap: the maximum primary-column width now derives from each label's `visibleWidth + 2`, with a minimum of 12; Pi still clamps to the available terminal width. No menu-specific override or new abstraction. - Validation: `extension/commands.test.ts` 89/89; `extension/controller-api.test.ts` 95/95; `npm run build` passed; `npm run check` 834 passed, 1 skipped, 0 failed; `git diff --check` passed. - Prettier ran once with temporary repository-root ignores; no schema/CHANGELOG churn remains. Live smoke: NOT RUN. ## External Review #4 resolution and validation - Shared contextual help is bounded and normalized to 160 display characters. - Manager/Chief supervision labels use a visible-width primary column (`visibleWidth(label) + 2`, minimum 12). The model picker is unchanged. - Validation: `extension/commands.test.ts` 89/89; `extension/controller-api.test.ts` 95/95; build passed; `npm run check` 834 passed, 1 skipped, 0 failed; `git diff --check` passed. Prettier ran once using a temporary ignore that preserved `docs/reference/agent-definition-schema.md` and `CHANGELOG.md`. Live TUI smoke: NOT RUN. --------- Co-authored-by: Jeffrey Boadi <25706638+boadij@users.noreply.github.com>
boadij
added a commit
that referenced
this pull request
Oct 6, 2026
## Summary - Make the exact current `ProjectAssignment` the durable authority edge for managed Leads. Manager reports and staff actions include only live exact assigned Leads; shared worktree scope alone grants no Manager authority. This intentionally removes ambient Manager control over unassigned Leads, which remain eligible for Chief supervision. - Automatically return each completed direct turn from an assigned Lead to the Manager role. When a herd run owns the work, its settlement remains the single automatic handoff. Results are nonterminal and routine progress is informational. - Add confirmed `/takeover` to remove the current assignment and pending project messages while preserving the Pi session and conversation, branch, worktree, running process, and Lead-owned Agents. Manual continuation of the exact assigned session remains managed until explicit takeover. - Serialize Manager `staff_stop` with the cross-process assignment lock so takeover cannot revoke authority during stop. Revalidate Manager lease and direct-report identity after `staff_inspect` evidence collection. - Derive and render `lead · managed` from current assignment state, without changing the Agent ownership breadcrumb. Update current documentation and add ADR 0017. - Clarify that `supervisor_message` queues nonblocking material coordination before the normal result boundary; it does not wait for a Manager reply or block project work. It is not for routine status or duplicate handoffs, and the example now describes a compatibility risk rather than a blocked Agent. ## Validation - Review #1 focused authority/takeover regressions: **3 passed**; `npm test -- extension/controller-api.test.ts`: **100 passed**. - Review #1 `npm run validate`: **848 passed, 1 skipped** (849 total); build and package audit passed. - Review #2 focused smoke-sequencing regression: **1 passed**; `npm test -- extension/smoke-runner.test.ts`: **60 passed**. - Review #2 `npm run validate`: **849 passed, 1 skipped** (850 total); build and package audit passed. - Final smoke-wait fix focused suite: `npm test -- extension/smoke-runner.test.ts` — **60 passed**; `npm run validate` — **849 passed, 1 skipped**, build and package audit passed (62 files). - Final formatting ran once with `prettier . --write`; two unrelated formatter-only files were restored without rerunning Prettier. Then `npm run check` passed (**849 passed, 1 skipped**, 850 total) and `git diff --check` passed. - Final documentation/guidance clarification: `npm run check` passed (**855 passed, 1 skipped**, 856 total); `git diff --check` passed. Prettier ran once; unrelated formatter-only changes were restored without a second run. ### Manager-recovery smoke history - The earlier pre-review feature-validation run passed once. - The first manager-recovery run after the review #1 authority fixes **failed (exit 1)** with repeated `Agent is already processing a prompt` errors after the scenario reached recovery and `staff_stop`. Review #2 identified an obsolete sequencing assumption: it prompted Manager after the Lead reply but before the automatically triggered project-message turn was known to be settled. - A separate diagnostic run before the sequencing correction **passed (exit 0)** because the automatic Manager turn happened to finish before the explicit prompt. That timing-dependent pass did not fix or reclassify the earlier failure. Its log remains at `.pi-herdsman/review1-smoke-diagnostic.log`. - Review #2 changed the smoke to wait for the exact project-message receipt matching branch, first Lead session, and context marker, then a settled Manager assistant turn, before sending its explicit review prompt. One invocation after this sequencing correction **passed (exit 0)** with cleanup passing. Its log remains at `.pi-herdsman/review2-smoke.log`. - A later run exposed missing model-wait classifications for the `direct-handoff-settled` and `graceful-close` stages. The review #3 run **failed (exit 1)** at `manager-recovery-graceful-close` with `SMOKE_STALLED` after about 15 seconds; evidence showed the expected READY handshakes had completed, but graceful close was not observed before the control stall timeout. Cleanup passed. This outcome remains recorded at `.pi-herdsman/review3-smoke.log` and was not reclassified. - The final smoke fix adds both `direct-handoff-settled` and `graceful-close` to the progress and stall-timeout model-wait lists. One final `npm run smoke -- manager-recovery` invocation **passed (exit 0)** through graceful close, resume/recovery, and retirement; cleanup passed. Full output: `.pi-herdsman/final-smoke.log`; validation report: `.pi-herdsman/final-smoke-validation.md`. Closes #278 Closes #277 --------- Co-authored-by: Jeffrey Boadi <25706638+boadij@users.noreply.github.com>
boadij
added a commit
that referenced
this pull request
Oct 8, 2026
Closes #294 ## Summary - Add three standalone, independently configurable Lead definitions: `flexible-lead`, `orchestrator-lead`, and `managed-lead`. Ordinary Leads select Flexible or Orchestrate; assigned project Leads use `managed-lead`. The definitions do not compose or inherit from one another. - Persist ordinary execution mode and its normal tool baseline separately from role/authority state. Project native Pi active-tool projection and structured prompt sections through the selected profile, while preserving mandatory Lead coordination tools. - Add the `/herdsman` Execution surface, `/lead` selector/direct command, and a future-session `defaultLeadExecution` setting. - Make successful `/takeover` release Manager assignment authority and transition the preserved session to ordinary Orchestrate using `orchestrator-lead`, after profile preflight. - Preserve the ordinary tool baseline when a restricted Flexible profile is selected, and restore branch-specific execution mode, baseline, tool projection, and prompt guidance during `/tree` navigation. - Reconcile cached Managed execution after successful assignment retirement while preserving Managed state when assignment/topology lookup fails. After takeover releases authority, recover safely from local execution-state persistence failure and attempt fail-closed tool projection without masking the original error. - Keep temporary usage fixtures under `PI_AGENT_ROOT`; lifecycle socket fixtures use platform-appropriate paths and close clients/server before cleanup. Manager authority remains established only by the existing exact `ProjectAssignment` semantics, and Agent ownership semantics are unchanged. ## External review follow-up - **External Review #1:** Preserve the ordinary Pi tool baseline when Flexible has a restrictive profile; restore branch-specific execution state and projection during tree navigation; reconcile execution after assignment retirement; and recover safely if takeover persistence fails after authority release. - **External Review #2:** Keep role reconciliation independent from malformed branch execution state and fail ordinary execution closed; reject ordinary mode changes when a Managed Lead's assignment scope is unavailable; and make ordinary mode activation transactional, restoring prior state/tools on failure or invalidating execution and failing closed if rollback also fails. ## Validation - Focused regression cases — **5 passed, 0 failed** (ordinary activation rollback, failed rollback fail-closed behavior, malformed selected branch, and Managed assignment-scope behavior in trusted/untrusted variants). - `npm test -- extension/commands.test.ts` — **109 passed, 0 failed**. - `npm run smoke -- manager-recovery` — **PASS** on the single latest run. - `npm run validate` — **PASS**; 874 passed, 0 failed, 1 skipped; build and package audit passed. - Independent review — no findings. - `prettier . --write` — completed; only the two changed files were touched. - `git diff --check` — **PASS**. - Manual TUI acceptance — **NOT RUN**. The following flows remain unverified manually: `/herdsman` Execution ordinary/managed flows, `/lead` mode changes, Settings default, Definitions actions, and takeover. --------- Co-authored-by: Jeffrey Boadi <25706638+boadij@users.noreply.github.com>
This was referenced Oct 9, 2026
boadij
added a commit
that referenced
this pull request
Oct 9, 2026
## Summary Closes #320 Replace frequent unchanged-fleet observation with event-driven invalidation and bounded authoritative reconciliation, reusing Herdsman's existing Herdr subscription and single-flight snapshot paths. - Subscribe to relevant Herdr lifecycle and presentation events. Presentation-only invalidations refresh the visible status without requesting Agent health scans. - Coalesce presentation-only invalidations to a 2-second minimum interval. Lifecycle and reconciliation notifications remain immediate; an immediate event cancels any pending presentation timer, as does abort. - Reconcile after subscription acknowledgement, on `events_lost`, and after reconnect. An unexpected close of an established subscription notifies reconciliation exactly once per outage; the subsequent close after `events_lost` does not duplicate that notification. Events remain invalidations, never snapshot authority. - Move ordinary Lead/delegating-Agent and leaf identity reconciliation to 10 seconds. Chief/Manager full UI supervision observation also reconciles every 10 seconds; the Manager widget redraws every 2 seconds from its existing snapshot for elapsed-time presentation. - Route local controller and project-assignment changes to active supervision UI. Headless sessions do not start UI-only periodic supervision observation. - Keep explicit/model-context refreshes and action-time identity, role, lease, ownership, and assignment verification fresh and authoritative. No new cache, scheduler, or dependency is introduced. ## Measurement (deterministic fake-executor call counts) These are test-harness counts of `api snapshot` calls, **not timing, real-Herdr, or fleet-performance measurements**: | Harness | Baseline | This branch | |---|---:|---:| | Leaf status, 10 captured 2-second callback ticks over virtual 20s | 10 calls | Inconclusive; no valid post-change count | | Delegating status, 10 captured 2-second callback ticks over virtual 20s | 10 calls | 2 calls for 2 captured 10-second callback ticks over virtual 20s | | Chief supervision cadence (separate cadence test) | — | 2 calls for 2 captured 10-second callback ticks | | Presentation burst through Herdr socket and managed-agent controller wiring | — | 8 simulated presentation notifications resulted in exactly 1 fake-executor `herdr api snapshot` refresh; the coalescing timer was advanced deterministically | The baseline Leaf and Delegating results used the audited v0.22.0 code. These probes establish scheduled/counted call behavior only; they do not establish elapsed-time speedup, production resource use, real Herdr behavior, or multi-Agent fleet scaling. The burst result is a simulated fake-executor count, not real Herdr or fleet timing. ## Validation and review - Review-fix focused tests: `extension/herdr.test.ts` — 71 tests passed; the extension-contract presentation burst integration test passed. - Earlier focused suites passed: `herdr.test.ts` (69), `extension-contract.test.ts` (49), `controller-lifecycle.test.ts` (59), `controller-api.test.ts` (100), `commands.test.ts` (111), `presentation.test.ts` (85), and `agent-runtime.test.ts` (62). - Final formatted full check: `npm run check` — 891 passed, 0 failed, 1 skipped. `git diff --check` passed. - Packaging: `npm run build` and `npm run package:audit` passed (65 package files audited). - Final review of the review-fix diff reported no findings. - No live smoke scenarios were run. - External Review #2 test cleanup: the event-watcher tests now advance presentation timers deterministically and use `t.waitFor` for reconnect assertions, removing repeated 2.1-second sleeps and the leftover reconnect timeout. Focused validation (`npm test -- extension/herdr.test.ts extension/extension-contract.test.ts`) passed: 121 tests, 0 failed, 0 skipped. - Final validation after the cleanup: `npm run check` — 891 passed, 0 failed, 1 skipped; `git diff --check` passed; `npm run build` passed; `npm run package:audit` passed (65 package files audited). ## Limitations and review status No real Herdr or scaled-fleet benchmark was run. Leaf's post-change call-count probe was inconclusive, and the 10-second reconciliation cadence bounds discovery of changes that do not produce a local notification or relevant Herdr event. PR #326 remains open and is handed back to the user for another external review; it has not been approved or merged. AI-assisted implementation; please review the code and test evidence as usual. - External Review #3 CI fix: the lifecycle watcher test now ignores a server-side `ECONNRESET` only when the test's abort signal is already set; all other socket errors are thrown. The assertions that exactly two connections occur and no callbacks are delivered remain intact. - Validation for this fix: `npm test -- extension/herdr.test.ts` — 71 passed; `npm run validate` — 891 passed, 0 failed, 1 skipped, including build and package audit (65 files); `git diff --check` passed. --------- Co-authored-by: Jeffrey Boadi <25706638+boadij@users.noreply.github.com>
This was referenced Oct 9, 2026
boadij
added a commit
that referenced
this pull request
Oct 10, 2026
## Summary Issue #333: stopping an assigned Lead could leave its managed Agent tree running when the Lead was in a different workspace. The Manager cleanup discovery/revalidation used its own workspace, Lead delegation was not synchronized with stop, and cleanup success was inferred from formatted text. - Discover and revalidate owned Agents across workspaces, constrained to the exact target Lead workspace and full Agent generation identity. - Revalidate cascade parents across workspaces and guard runtime-cache invalidation by workspace so a foreign Agent with a duplicate label cannot invalidate local tracking. - Reuse the delegation lifecycle lock across Lead delegate/continue, Manager stop, and confirmed Lead stop-all. Manager stop keeps the assignment when descendant cleanup or Lead stop verification fails. - Return structured cleanup completion and verify the final owned-Agent inventory before stopping the Lead. No assignment, worktree, branch, or saved session retirement is performed. - Add cross-workspace duplicate-label, partial-cleanup, busy-lock, gated-startup, and stop-race regressions. ## Validation Validation on the preceding implementation head `60b435a4df34ed419570bddab9c2f280c7c37637`: - `npm run validate` — passed: 942 tests passed, 1 skipped; build passed; package audit passed (68 files). - `git diff --check` — passed. - `npm test -- extension/controller-api.test.ts` — passed (114/114). - `npm test -- extension/commands.test.ts` — passed (116/116) on the preceding implementation head. - `npm test -- extension/controller-lifecycle.test.ts` — passed (68/68). A prior test failure was traced to the new regression test statically importing `agent-controller.ts` before support test-module mocks were registered. The corrected test defers that import until after mock registration. For comparison, isolated baseline `b4ed98e` passed the pre-correction controller-api suite (113/113). The final validation above includes the correction. Live smoke was not rerun; earlier `manager-recovery` smoke covered ordinary stop/resume, not active recursive descendant cleanup. No force-push was made. Fixes #333 ### External Review #2 follow-up Final PR head: `6f231555398f5416bc557d2040e8a958778ffde7`. - Locked empty stop-all revalidation now uses the authoritative Lead Agent snapshot and refuses success when mailbox-state issues remain; newly visible Agents still require reopening the confirmation flow. - Added a regression proving unresolved mailbox state with zero visible Agents produces no success summary and starts no cleanup. The busy-delegation-lock regression remains in place. - `npm test -- extension/commands.test.ts` — passed (117/117). - `npm run validate` — passed (943 passed, 1 skipped; build and package audit passed, 68 files). - `npm run check` — passed (943 passed, 1 skipped). - `git diff --check` — passed. --------- Co-authored-by: Jeffrey Boadi <25706638+boadij@users.noreply.github.com>
This was referenced Oct 10, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
🤖 I have created a release beep boop
0.1.4 (2026-09-07)
CI
Other Changes
This PR was generated with Release Please. See documentation.