Agent orchestration v1: three-party loops, agentic settings, spawn knobs, mission-control fixes (L8-L18, mcp-v3.0.0rc3)#100
Merged
Conversation
…ev/flows Orchestration L0 (260703-L0): FlowTab becomes a pure segment renderer with a radiogroup model nav; content externalized to the new flowModels.ts registry carrying 8 static models (build-job, frame, designer, orchestrator, manager, worker, reviewer, comms) that encode the orchestration series' agreed invariants. Mounted dev-only at /dev/flows (initialModel + ?model= deep link); FlowTab.test.tsx asserts nav, fallback, per-model render census, and the invariant texts verbatim. Shipped dashboard bundle rebuilt and synced. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
…s variants, report templates Orchestration L1 (260703-L1): the behavior layer as doctrine. SKILL.md carries the frame (context -> job selection -> housed execution -> wrap-up), the never-self-spawning entry rule, the coordination-leaf convention, the escalation ladder, the two adversarial seams, gate-delegation doctrine, the knob block + per-harness variant resolution, and the settings orchestration schema. Jobs: designer / orchestrator (sole spirit-test carrier) / manager / worker / adversarial-reviewer, plus claude-code exemplars for orchestrator and worker. Six report templates. Synced to all starter packages via scripts/sync-skills.py; runtime code (spawn tool, comms, gate enforcement, config parsing) stays deferred to L2-L5 as referenced contracts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
…existing session primitives
Orchestration L2 (260703-L2): orchestrators spawn managers and managers spawn
workers without dashboard clicks. A shared serving/terminal_opener.py backs
both the dashboard opener route and the tool (leaf claim before spawn,
env-seeded tmux ensure, catalog upsert); serving/terminal_paste.py adds
server-side echo-confirmed paste behind POST /api/terminal/{session}/paste;
knob passthrough rides tmux -e env seeding (AR_SPAWN_MODEL/AR_SPAWN_EFFORT,
CLI-flag mapping deferred to the role-knob layer); spawned-by provenance lands
on the catalog row; 409 leaf-taken is surfaced, never overridden. Strict
pydantic response contract; tests incl. test_spawn_agent_session.py; full
quality wrapper green (1411 passed, 0 CRAP threshold breaches).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
… turn artifacts, nudges Orchestration L3 (260703-L3): inbox rows gain sender/recipient role addressing, message kinds, artifact paths, and hosted delivery state; serving/ inbox_delivery.py pushes queued messages into hosted sessions over the L2 paste seam while durable polling stays the source of truth. Fixes the L2 paste false-positive during harness boot (retry over the boot deadline; delivered only on a real pasted-draft/chip echo — regression-tested, the defect the first live spawn dispatch exposed). Adds orchestration artifact helpers (turn-report paths, master-handover rendering, one-rung escalation packets) and the rate-limited nudge policy with observer orchestration.nudge events + orchestration_nudge_manager. Skill worker/template conventions updated in the canonical tree + all synced packages. Full suite 1418 passed; wrapper exit 0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
… with attributed decisions Orchestration L4 (260703-L4): controlplane/gate_policy.py defines the validated delegation schema — all-human default, strictly opt-in, human-pinned kinds (integration-approval, push-approval, cleanup-approval) never delegable, and silently-weakening policies rejected at validation. evaluate_closeout_gate generalizes into the kind-generic evaluate_gate(policy=...) resolver; gate records/payloads carry decidedBy/decidedVia/decidingRole + append-only evidenceRefs; owner self-approval is server-rejected before append; reviewer- verdict requirements are enforceable per policy; projection exposes the delegated-decision fields (GateResponder render polish deferred). Delegation is additive: human approvals stay valid everywhere. Wrapper exit 0 post-merge with the landed L3 comms layer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
…atch order, conflict modes Orchestration L5 (260703-L5): the l-02 skill carries the full branch topology — super off main, masters off super, leaves off masters, C-11 at every edge — plus dependency-ordered manager dispatch, parallel-master reconcile over a moved super base, the orchestrator master-to-super integration worktree procedure (now also in jobs/orchestrator.md), the two conflict-resolution modes, leaf-move decision logs, the memory-ledger invariant, and the final super-to-main PR landing tail. Master finalize/archive + parallel-master reconcile tooling explicitly SEQUENCED to 260703_task-doc-tooling-repair T8/T9 (the archived 260630 backlog's new home) — manual orchestrator duties until then. Doctrine-only: no Python changes; skills synced to all targets. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
…nce attachment procedure Orchestration L6 (260703-L6): jobs/adversarial-reviewer.md gains the MASTER-EXIT and SUPER-EXIT rubrics (scope packets, completion vs task docs, tools.md quality, onboarding-vs-code, refute-or-confirm posture, decomposable fix-leaf rules); templates/verdict.md splits into the two seam variants with ranked findings, pass/pass-with-notes/block, and the exact reviewer-verdict gate evidence ref shape; SKILL.md documents the seam procedures (manager spawns at master-exit, orchestrator at super-exit, verdicts land in series notes/reports/ and attach via the L4 evidenceRefs). No Python glue added — the existing GateEvidenceRef / require_reviewer_verdict contract suffices. Doctrine-only; skills synced to all targets. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
The master change-set read 0 for every in-flight series: it diffed the master base against the source-branch tip (main) while the series commits live on the work branch until the final PR lands. A shared _series_tip resolver now prefers the live work branch and falls back to the source branch once the series has landed; master_changeset and master_file_diff use the same resolved tip so counters and file view cannot drift. Regression tests cover the in-flight and landed shapes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
Lifecycle and job are one entity: each agent type runs its own self-contained lifecycle. One skill folder replaces both l-01-session-job-lifecycle and l-02-agent-orchestration: SKILL.md is a three-condition router (AR_SPAWN_ROLE -> role brief -> orchestrator) plus the minimal frame; the orchestrator lifecycle absorbs the old session-job spine with solo work as the degenerate portfolio and is the topology's single home; the worker is rewritten positively (no l-01 references, closeout/integration stripped to the owning seat, brief as session start); lenses.md replaces job-variants; templates gain worker-brief.md (the institutionalized dispatch packet) and the relocated deep-research-report.md. Code edges: orchestration_artifacts template root, next_step rundown wording, abandon/reducer comments. Four c/w skill references updated; all mirrors synced and pruned. Design record: tasks/agents-remember/260703_agent-orchestration/notes/ design-lifecycle-convergence.md (12 findings, merge map, S4 rulings). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
The reopened convergence pass gets the orchestrator's shape right: an event loop over durable portfolio state (opening move: trust checkpoint + portfolio orientation) routing into Design (the designer hat, pulled inline — the designer is a hat, not a seat), Portfolio (bulwark + reshape + the planner master; ordered-list renumbering doctrine), and Orchestrate (super-branch INTENT creates a branch only; the orchestrator worktree exists per master-to-super integration edge). The invariant ladder — approved task doc, then branch, then worktree only where something is built — replaces the transplanted worktree-first ordering; chat is never a build route (260628 T7); the failed-deliverable rule (task_reopen + reshape, never redo siblings) and the hat-collapse rule (manager hat in flat runs, hands-on at session scale) are encoded where the seats read them. Design record: tasks/agents-remember/260703_agent-orchestration/notes/ design-orchestrator-routes.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
…roles Harness abilities are capabilities, not vendor features: the two .claude-code.md overlays are deleted and their content generalized into capability-conditional sections of the portable files — the orchestrator gains "Sub-Agent Fan-Out (any harness that has it)" (durable reports + compact summaries, AR mutations stay in the main loop, concurrency-capped, sequential fallback without the ability) and the worker gains the read/search-scoped equivalent with the never-delegated turn report. SKILL.md drops the per-harness variant layer: resolution is role-file defaults < settings.json orchestration block; harness preference is deployment configuration, and no per-harness role files exist by design. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
The 17-finding review (notes/reports/260705-lifecycle-suite-adversarial- review.md) lands per the developer's rulings. AR-1: the master-exit handover gate is the new delegable GateKind master-handover-approval — the manager RAISES it with the reviewer verdict attached, the ORCHESTRATOR decides (human review concentrates at the super gate); identity mechanics documented; requireReviewerVerdictAtSeams is WIRED (apply_seam_verdict_requirement binds delegated seam rules to verdict evidence — AR-5). The reviewer file is roles/reviewer.md (server vocabulary; spawn value stated everywhere — AR-3); the router decides its edge cases in writing and templates gain manager-brief.md (AR-12); the manager loop names the human-pinned kinds (AR-4); the lenses chat remnant, dead variant rungs, retired vocabulary, dangling refs, and the six-signals phrasing are swept (AR-2/7/9/15/16); hat-collapse gains the gate-reversion clause (AR-6); the hat/seat exception is stated (AR-8); takeover has an owner (AR-10); the rundown speaks the event loop (AR-13). The FlowTab canvas is redrawn to the converged doctrine — ROUTER in, FRAME/BUILD-JOB out, all role models on the ruled semantics (AR-14) — visuals now ride every doctrine change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
… enforcement Review round 2 (notes/reports/260705-lifecycle-suite-adversarial- review-2.md) proved the ruled seam deadlocked as-built (AR2-1) and had no enforcement consumer (AR2-2). This cycle lands the channel: lifecycle_gate gains wait=false — raise-and-continue, allowed only for kinds the active policy delegates — returning the gateId the handover packet now carries (new packet row); gate_decide resolves a bare gate id across lifecycles (GateStore.find) so the orchestrator decides by the packet-carried id without ever handling a lifecycle id, and refuses cli-attributed decisions on delegated kinds (fail-loud instead of the policy bypass); worktree_integrate refuses while a master-handover- approval gate exists undecided or policy-invalid (mirror of the closeout consumer); seven SeamChannelTests exercise the flow end-to-end. Doctrine follows: manager.md states the exact raise-then- packet channel, orchestrator.md names the decide call and gains the Profile check (takeover) the AR-10 pointers referenced, the router/ capability sections state the spawn-as-fan-out backdoor (the framework is the harness-independent fan-out — the DBMS principle), reviewer/ designer knob rows gain the inbox, worker-brief carries the canonical ROLE BRIEF header, settings-json.md documents the extended named policy, and the remaining round-2 nits are swept. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SjabEc4oGWv5VB1x9eyYtY
…ion (AR3-1..AR3-6) worktree_integrate_tool now passes the configured gate policy and the integrate guard addresses master-handover-approval gates by MASTER identity (enclosure = master task name) folded cross-lifecycle via GateStore.all_current(), closing the inert-consumer MAJOR from adversarial review 3. wait=false raises validate-then-mutate and are restricted to seam gate kinds; gate_list defaults to the ambient lifecycle; manager doctrine carries enclosure on the raise plus the all-human conditional; c-09 names the handover-gate-blocked state; canvas seam nodes, manager-brief planner-master slot, and handover-packet manager row ride along.
A master-handover-approval raise-and-continue now refuses a missing enclosure before any store mutation, and integrate surfaces open-but- unmatched handover gates as a spelling-check warning while staying additive when gateless. Dry-run evaluates the seam guard and reports would-block without persisting contract state. The canvas raise node names the enclosure address, the panels overview claims exactly what the canvas draws, and the l-01 template registry lists all nine templates.
GET-only /api/notes/{list,read} with repo allow-listing, single-segment
master validation, and realpath confinement (registered before the greedy
static mount); the TaskNotes panel lists a master's notes and renders
opened notes through the shared Markdown treatment; task-doc references to
notes files resolve into openable links. Review follow-up L9R-1: ValueError
from Path.resolve() (null-byte input) maps to 400 bad-path in both the
notes and files status mappers, with regression tests in both suites.
EnclosureNode gains codeWorktreeExists/memoryWorktreeExists, stat'ed at snapshot time; the Hangar and the tasks list filter on physical existence via the shared hasLiveWorktree rule (a reopened leaf stays hidden until worktree_start recreates its worktrees), and each leaf renders exactly one row (one-row-per-enclosureId with deterministic lastEventTs annotation). Review follow-ups: the rich sim fixture builder materializes the worktree dirs its live contracts record (regression suite added), and the anchor fallback annotation is order-independent.
The l-01-agent-lifecycles skill is the single public truth: the dead orient→ground→frame→decide axis, retired skill names, l-02 contact-point language, and the chat-build intake are swept from living docs, templates, and ALL TEN first-action surfaces (four hooks + Gemini/Hermes/OpenClaw starters + the Pi extension + Cursor rules + Copilot instructions, each byte-consistent with its package hook so the install-doc quotes stay true). Chat is never a build route: single-session work rides a thin w-02 doc (c-09 intake, w-02 prose, FAQ). Canvas residual: the designer node's stale phase word. server.py front-half docstring joins the sweep.
…op law The owner→builder→reviewer loop is a first-class lifecycle feature at every level that owns work. NEW: roles/strategist.md (spawn-first sprint planner, MANDATORY before any orchestrated run: eight-phase method — inventory, two-sided touch surfaces incl. greenfield-by-declaration, cgc/grepai dependency edges with evidence, blast-radius register feeding loop-tier scoring, coherence sweep, topological ordering — producing the ORCHESTRATION TASK = the sprint scope); templates/orchestration-task.md (shown-work schema); five criteria catalogs under criteria/ with the promotion ratchet (candidate at 1 catch, standing at ≥2, gate when mechanizable). Loop law in SKILL.md: complexity tiers (direct/builder-verified/full), hard 3-full-round cap with delta-verifies exempt, the CONVERGENCE rule (a non-shrinking round escalates regardless of count), same-builder/same-reviewer reuse, quo-vadis. Orchestrator super-exit rewritten per the developer ruling: integrations are orchestrator-delegated; the developer reviews ONCE at the super PR gate with a REVIEWABLE ENVIRONMENT, visible-behavior-first, demo notes obligatory. Canvas gains the strategist model (8); server roles gain strategist.
New kernel/agentic_settings.py: per-use read of the GLOBAL ar-coordination/system/settings.json merged with <repo>/system/settings.json (leaf-key deep merge, arrays replace, orchestration.* unknown keys fail loud naming the offending file). The L12 knob schema finally parses: loops, role knobs, concurrency caps, spawn harness preference (registry-validated; explicit > repo-local > global > detection, refusals name their source). gateDelegation migrates to the global file (boot snapshot; authority-file value = one-cycle warned fallback; repo-local values REFUSED — gate posture is workspace-wide, and validate-then-ignore would be fail-open). The old implicit provider-settings fallback (fail-open: empty config on missing file) is removed; memorySettingsIncludes dead plumbing removed; contextProviders documented as a reserved future family. c-13 gains the agentic-settings interview; runtime_install seeds the global file copy-if-missing; settings-json.md is the single schema reference.
The developer-picked V4 treatment lands as one insignia language: a shared RankBadge chevron component (gold three-chevron under a command pip for the orchestration tier, purple two-chevron for management; six new OKLCH tier tokens) worn by the tasks tab AND the chats pane. Tasks tab: orchestration rows (master-kind docs with the new additive `orchestrates` field) render the folded-corner/ghost-wash treatment above their commanded masters with N-depth nesting and a cycle guard; flat runs render exactly as before (D3). Chats pane: the G1 command tree — a sprint deck of command chats grouped by the new persisted spawnRole provenance, master groups by leaf claim (grouping-always, owner-ratified), and a landed archive roll-up. task_doc, projection, and the sessions wire carry the two new fields additively.
The long-session OOM's real driver (measured, not assumed): five volatile now-relative age fields defeated the SSE diff every tick, re-emitting the whole analytics payload (~780 KB/s idle). The delta stream now diffs stable forms (VOLATILE_AGE_FIELDS stripped; served ages advance locally from arrival anchors), /api/state honors If-None-Match with a content ETag, and the client store is identity-preserving and change-gated — idle wire traffic is zero bytes and getState() identity is stable across ticks (500-tick guard test). tokenSeries is decimated serving-side (512 cap, newest 256 exact; the log untouched). A reflection guard classifies every *Seconds projection field as volatile or content so one-sided lockstep drift fails loudly. The dashboard shows its serving build (commit + boot time) so a stale server is visible at a glance.
…egistry
AR_SPAWN_MODEL/EFFORT now map onto the harness argv through the registry
(claude: --model/--effort; effort value `ultracode` delivers as a
"/effort ultracode" session paste submitted BEFORE the role brief), with
dispatch-time refusals naming the harness and both value sets — the CLI's
warn-then-silently-degrade is unreachable, and the opener re-validates as
defense in depth. The registry becomes good defaults, not a wall: the new
orchestration.harnesses settings family adds hermes-style harnesses or
pre-customizes builtins with user command arrays; vocab-less user
harnesses refuse the effort knob with declare-or-launchArgs guidance.
Free-form vehicles: launchArgs (verbatim argv), sessionCommands (own
submitted pastes), promptKeywords (brief's first line) — never validated,
always in provenance. rolesPerLevel {leaf|master|portfolio} deep-merges
over flat role defaults; spawn gains a level input recorded in
provenance. docs/reference/harnesses.md is the parameters manual with a
worked add-hermes example. Scratch terminals never load the registry.
NotesReaderViewer takeover: rail = the master's notes tree from /api/notes/list (ChangeSetViewer idiom, highlight follows selection), content pane = DualPane reuse (markdown/text/binary all trace DualPane's own branches — no second bespoke reader). TaskNotes reduced to the entry surface; DetailPanel threads onOpenNotes (11 sites); CockpitShell holds the takeover with file-viewer hidden-not-unmounted persistence. Supplements: gateHint ask-chip fallback removed (+ regression test pinning its absence); wait-loop-era remnant sweep — 3 vestigial surfaces fixed, remaining ask surfaces verdicted legitimate, 2 follow-ups named (GateResponderText dead branches, summarizeBlocked wording). Checks: dashboard 45 files / 433 tests passed; tsc, build, sync-dashboard --check, quality wrapper all exit 0. Adversarial review PASS-WITH-NOTES (260703-L17-adversarial-review.md).
…chet) code-seam CS-4 — reused-primitive affordance parity (catch: L17R-2, the DualPane markdown path dropping the truncation banner). report-verification RV-4 — decision-log completeness for scope-expanding supplements (catch: L17R-1, supplement 2 recorded only in the builder report). Both single-catch candidates per the ratchet; accepted by the loop owner in the L17 decision log. Mirrors synced via scripts/sync-skills.py. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Snapshot of the agent-orchestration series before the end-to-end orchestrated pilot: three-party loops (L12), settings unification (L13), visual command hierarchy (L14), flat-memory dashboard serving (L15), spawn knob application + open harness registry (L16), notes reader v2 (L17). The pilot runs against this packaged runtime. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
F1 (HIGH, closes F-N): TaskDocResponse declares removedSubtask/
deletedFiles/wouldDeleteFiles — remove_subtask no longer reports a false
error after destructive success. F2 (L17R-2): truncated-markdown 2-MiB
banner in NotesReaderViewer. F3 (F-A seam): inbox delivered/unconfirmed
regression pin, canonical-invocation-sensitive via the sys.path idiom.
F4: effortSessionCommand template validation in the settings kernel
(only {value}; loud refusal names file + harness). F5: decode_capped
UTF-8 boundary backscan wired into notes/files serving (+ third site).
F6 (developer-ruled): JSON null at any orchestration family key refuses
loudly ("remove the key to inherit the global value"). F7 (F-R):
worktree_start 'reconciliation' recovery implemented — ledger mapping
recorded the closeout way; recovery blocks advertise only consumable
choices.
Ratchet: RV-2 + RV-4 promoted to standing (second catches at this
leaf's review); RV-5 + CS-5 seeded as candidates; mirrors synced.
Checks: mcp pytest 1628 passed / 3 skipped / 192 subtests; dashboard
435 passed; ruff, pyright, typecheck, build, sync-dashboard --check,
quality wrapper all clean. Adversarial review PASS-WITH-NOTES +
same-reviewer DELTA-PASS (260703-L18-adversarial-review.md).
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 575a9a44b7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+1183
to
+1185
| write_ledger(contract.memory_repo_path / "memory.md", updated) | ||
| require_git(contract.memory_repo_path, ["add", "memory.md"]) | ||
| commit_if_dirty( |
There was a problem hiding this comment.
When the memory repo is checked out on a branch other than contract.memory_source_branch (a case this module already accounts for via _memory_base_for_source), this writes and commits the reconciliation mapping to whatever branch is currently checked out. The rest of worktree_start still creates/uses the memory worktree from contract.memory_source_branch, so if that branch already exists the new mapping is absent from the worktree and the start reports compatible even though the branch it starts from remains unmapped. Commit the reconciliation to the intended memory source branch (or check it out/update it explicitly) before proceeding.
Useful? React with 👍 / 👎.
Comment on lines
+230
to
+232
| launch_args=override.launch_args or base.launch_args, | ||
| prompt_keywords=override.prompt_keywords or base.prompt_keywords, | ||
| session_commands=override.session_commands or base.session_commands, |
There was a problem hiding this comment.
When a flat role default sets launchArgs, promptKeywords, or sessionCommands and a rolesPerLevel override intentionally sets that array to [], _parse_roles produces an empty tuple but this or treats it as absent and inherits the base value. That violates the documented array-replace semantics and causes unwanted launch args, prompt keywords, or pre-brief commands to be injected for the level that tried to clear them.
Useful? React with 👍 / 👎.
Codex P1: _reconcile_missing_mapping refuses (naming both branches) when the official memory repo is checked out on a branch other than the memory source branch — the mapping commit must land on the branch the worktree is created from, never on whatever happens to be checked out. Codex P2: an empty free-form list (launchArgs/promptKeywords/ sessionCommands) is refused with inherit guidance — RoleKnobs cannot distinguish absent from cleared (empty tuple = not configured), so [] meant to clear a flat default would silently inherit it through the field-wise or-merge; same silent-degrade shape as a null family key (L18 finding 6), same refusal remedy. Both with failing-first regressions (166 tests green in the two suites). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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.
What this lands
The 260703_agent-orchestration series — the v1 multi-agent orchestration system, built and reviewed through the very loops it defines (28 commits, ~39k insertions):
system/settings.json, leaf-key deep merge, fail-loud on unknown keys, gateDelegation boot authority.launchArgs/sessionCommands/promptKeywords), unknown harnesses teachable via settings instead of crashing;docs/reference/harnesses.mdmanual.remove_subtaskfalse-error, theworktree_startreconciliation recovery, UTF-8 boundary truncation).Verification
Every leaf landed through a full loop: builder report → adversarial review (opus) → owner folds → delta-verify, with closeout body gates on onboarding. Verdicts + reports live in
ar-coordinationtasks/agents-remember/260703_agent-orchestration/notes/reports/. Final state: mcp pytest 1628 passed / 3 skipped; dashboard 435 passed; ruff, pyright, build, sync checks clean (also enforced by the pre-push hook on this branch).Release tag:
mcp-v3.0.0rc3on the head commit. Paired memory-repo branch:ar/260703_agent-orchestrationin the memory repo (carryover follows the merge).Post-merge plan: memory carryover → publish rc3 → flip workspace MCP configs to the uvx install path → nine-master orchestrated stress test on the released runtime.
🤖 Generated with Claude Code