orchestra/AUDIT.md

Orchestra — spec conformance audit & remediation status

Audited 2026-07-27 against orchestra-spec (1).md at commit 325c684. Method: read every non-test file in internal/ and cmd/, traced each spec section to its call site, and checked whether the live path (main.go → router → coordinator → adapter) actually reaches it. This file is the single merged record — a prior split between AUDIT.md (audit + plan) and progress.md (chronological session log) has been flattened here; treat any older chronological git history of either file as session notes, not ground truth.

Ground truth rule for future sessions: this repo has a documented history of code that looks wired but isn't (packages with tests that pass in isolation while the live call path silently no-ops). Before trusting any "done"/"fixed" claim below, check the actual call site, not just this document.

---

Current verdict (as of 2026-07-28)

The substrate (Layer 1) is solid. Layers 2–4 were originally shaped-but-not- wired; most of the blocking defects below are now closed, verified by reading the current code (not just by trusting this file) and by targeted tests. Three things are not yet proven:

1. A real end-to-end unattended run. The live attempts of 2026-07-28 surfaced B12 through B17. All are now closed in code (B17 on the worker path only), but B13 through B17 were each found live and none has been re-verified live — occupancy/rotation/handoff/completion still haven't been exercised end-to-end against a session guaranteed to exist. Treat "closed" here as "the code no longer contains the defect," not as operational proof.
2. The operator surface was unauthenticated in the live deployment. Fixed in code (B18: mandatory Web token, cookie sessions for the browser), but the fix is a breaking config change — the service will not start until .orchestra-config/orchestra.env sets ORCHESTRA_WEB_TOKEN. See B18 and the command-channel section.
3. Cross-machine (federation) correctness. Deliberately deferred — see "The federation fork" below.

go build ./..., go vet ./..., and go test ./... all pass.

Follow-up live verification (2026-07-28)

The OpenCode launch path has now been exercised against workpc herdr through an isolated test-e2e worktree. The verified outcomes are deliberately narrow:

• Initial prompts use a bounded retry only after an explicit JSON-RPC rejection (herdr's short post-start readiness window). Transport timeouts and disconnects are never replayed. A pane is inspected first; blocked or permission-dialog panes are refused.
• Herdr accepts task-scoped unique names (oc-<task-id>); active sessions are visible by that name in agent.list. Prompt routing uses that name, while pane reads remain pane-scoped.
• The initial launch message carries the task title and description. This is necessary under Design A because homesrv cannot write TASK.md inside a workpc checkout. A live OpenCode run received the full instruction and, after explicit operator approval of two file edits, created the requested marker and report without touching other files.
• The initial launch does not wait for the agent to become idle. Waiting converted an ordinary long-running first turn into a false TaskBlocked; a fresh run now remains TaskLeased while OpenCode is paused at its normal permission boundary.

Deployment state observed live (2026-07-29)

Checked against the running system, not the docs:

• The deployment moved from orchestra.service to Docker Compose (/home/kami/docker-apps/orchestra-web-ui). orchestra-api and orchestra-web-ui are both up; /healthz and the UI answer 200. The stopped systemd unit is the retired predecessor, not an outage — see the deployment section of CLAUDE.md.
• The running image predates the last two commits (built 2026-07-28 21:28; 0b4b52a "Require a token for the web UI" landed 23:15). So B18's auth fix is not in the live container. Rebuild before drawing any conclusion about the live surface's auth behaviour.
• All six herdrs are unreachable — workpc 192.168.1.105:9245-7 refuses, homesrv 192.168.1.104:9245-7 times out (filtered). Probed directly, since a herdr that connects logs nothing. Nothing can be leased; every task the API serves is history. This is the sole remaining blocker on live proof for B13–B17.
• GET /v1/tasks has no authz check at all (cmd/orchestra/main.go:286 returns before any gate). B18 gated the mutation and approval controls, not the read path. Accepted for now: the 0.0.0.0:9145 bind is deliberate and ufw restricts the port to a single other LAN machine, so the read path's trust boundary is the LAN, not the host. Revisit if that bind is ever widened.
• ntfy delivery still fails 403 Forbidden on every send (last observed 2026-07-28) — a wrong or expired credential, separate from S12's ORCHESTRA_NTFY_SURFACE_TOKEN.

This does not complete the cross-machine continuity design. clients/herdr- bridge.go is only a byte proxy; it is not a worker. The worker-side process described below remains required before a remote checkout can author TASK.md, derive a canonical handoff anchor, scratch-commit WIP, or safely release it. Until then, homesrv must not rotate or clean up a non-local worktree.

Spec layer	State
L1 substrate (§3, §4)	Built and correct in the main path.
L2 harness (§5)	Occupancy, rotation, and completion all wired and reachable. B12 fixed and confirmed live; B13–B16 fixed in code but not re-verified live. Unique per-task agent names (B16) and ambiguity-safe prompting (B14) are the substantive changes.
L3 continuity (§6)	Handoff schema, pickup validation, scratch branches, TASK.md are all wired into the live path (Phase 4 complete).
L4 surfaces (§7)	Brief/standup/delivery real; quota has a post-hoc producer only (no live push feed yet); B8's authz bypass closed, but the new web UI reintroduces an equivalent one (B18).

---

Blocking defects

B14, B15 and B16 are now closed by code and covered by tests; none has been re-verified against a live cross-machine run. B17 is closed on the worker path only and remains open for Design A. B18 (the unauthenticated web surface) is closed in code but requires an env change before the service will start — see B18 and "What's next".

Nothing is currently blocking in the sense B12/B13 were. B19–B21 and S12–S13 — the correctness and hygiene gaps found while implementing B18 — are now fixed in code and covered by tests (2026-07-28); each is marked closed in its own section below. The flaky router test is also fixed, and the flake was in the test, not in assignment.

The real remaining risk is entirely evidential now: B13 through B17 were each found live and none has been re-verified live, and none of the 2026-07-28 fixes has run on the deployed instance either — the service is stopped, /usr/local/bin/orchestra predates all of them, and installing a new binary or editing /etc/orchestra/orchestra.env needs privileges this sandbox does not have.

B18 — the web UI surface is unauthenticated (found by audit 2026-07-28, uncommitted working tree) — closed (code fix; requires an env change before restart)

cmd/orchestra/main.go:205 mounts the browser read/write surface with no handler-local credential check:

mux.Handle("/v1/ui/", ui.Server{Store: s, Workers: workers, Coordinator: coordinator, ...}.Handler())

Unlike the worker/harness endpoints on the same mux (main.go:342, 372, 467, 821, 840), it performs no check of its own and stamps its events Surface: string(authz.Web) (internal/ui/ui.go:222, 337, 368).

It is not, however, ungated in principle. authz.HTTP wraps the entire mux (main.go:1167) and defaults a request with no X-Orchestra-Surface header to Web, so /v1/ui/ is covered by ORCHESTRA_WEB_TOKEN. The defect is that this gate is opt-in — authz.HTTP's own comment says "authentication is optional for local development," and expected != "" is the only thing standing between the LAN and the surface.

The live deployment has opted out. .orchestra-config/orchestra.env sets no ORCHESTRA_*_TOKEN at all, and the server binds ":"+port (all interfaces, 9145). So on the deployed homesrv instance this surface is in fact reachable unauthenticated from the LAN.

Consequence. Unauthenticated reachable actions include TaskCreated, and per-task release, block, complete and handoff (ui.go:345-361). Most seriously, grant_approval / deny_approval (ui.go:301) terminate in pane.send_text against a live pane — so an unauthenticated caller can inject keystrokes into a running agent session on workpc. B14 was filed because Orchestra could accidentally advance a permission dialog; this surface allows it deliberately, from the network.

Why simply setting ORCHESTRA_WEB_TOKEN does not fix it. mux.Handle("/", webui.Handler()) serves the SPA from the same listener, behind the same middleware. A browser cannot attach an Authorization header to a top-level document load, so setting the token returns 401 for index.html and the web UI stops loading entirely. Bearer-token auth and a browser-served SPA are incompatible as currently wired; that is likely why no token is set.

Required fix (design decision, not a one-liner). The browser needs a credential it can actually present on a document load — a login endpoint that exchanges the Web token for an HttpOnly, SameSite=Strict session cookie, with authz.HTTP accepting either that cookie or a bearer token for the Web surface. Alongside it: make the token mandatory rather than opt-in whenever the UI is mounted (refuse to start, rather than silently serving an open control plane), and bind to loopback by default so exposure is a deliberate configuration act. Approval commands in particular must not be reachable without it.

Fix (2026-07-28):

1. The token is now mandatory (cmd/orchestra/main.go): startup calls log.Fatal if ORCHESTRA_WEB_TOKEN is empty. Serving the control plane openly is no longer something a missing env var can cause silently.
2. authz.Sessions (internal/authz/authz.go) issues 32 random bytes as a session value and stores only its SHA-256, so a leaked snapshot of the map yields nothing usable. Sessions expire (12h default) and are swept on each issue.
3. POST /v1/ui/session verifies the Web token in constant time and sets the value as an HttpOnly, SameSite=Strict, Secure cookie. Secure is dropped only if ORCHESTRA_UI_INSECURE_COOKIE is set, which is required to reach the UI over plain HTTP.
4. authz.HTTPWithSessions accepts that cookie in place of the bearer token, and only for the Web surface; the bearer comparison is also now constant-time. Two paths are exempt from the gate, both necessarily: the login endpoint (which performs its own check — gating it would make login unreachable) and non-/v1/ GET/HEAD, i.e. the SPA shell and its static assets, which are not secrets. Every /v1/ control path stays gated.

TestWebSessionCookieGatesControlPathsOnly asserts the whole shape: an uncredentialed control path is 401, the shell and login endpoint are reachable, a valid cookie authorizes, a forged one does not, and a cookie presented on the TUI surface is rejected. TestSessionExpires covers TTL.

Operational precondition — this is a breaking config change. .orchestra-config/orchestra.env currently sets no tokens, so orchestra.service will refuse to start until ORCHESTRA_WEB_TOKEN is added (plus ORCHESTRA_UI_INSECURE_COOKIE=1 if the UI is served over plain HTTP). Setting a Web token also newly gates the other /v1/ surfaces that default to Web — any existing unauthenticated client of this instance will begin getting 401s and needs the token too. Do not restart the service without making that env change first.

Still open: the listener binds all interfaces (":"+port); loopback-by- default was considered and deliberately not taken, so exposure remains a network-level concern.

B14 — agent.prompt can duplicate a task prompt after an ambiguous wait timeout (found live 2026-07-28) — closed (code fix; not re-verified live)

Live Claude E2E task 06FTF8CPH3K3DPN6XQA7G3WRQ8 (pane wM:p1) received the same initial Begin Orchestra task ... prompt twice. Claude completed a first turn inspecting the repository and asking for task context, then began working on the identical prompt again after the turn stopped. This was observed directly in the pane transcript; it is not merely a UI replay.

Root cause: CLIAdapter.Lease calls Client.Prompt with wait.until=idle and a 60-second wait. Client.Call, however, applies the client's default 10-second connection deadline when the caller context has no deadline. Once that local read deadline expires, Prompt treats the result as any other transient failure and retries agent.prompt with the same payload for up to bootRetryWindow. Herdr may already have accepted the first request, so this is an ambiguous outcome, not a safe retry.

Consequence: an agent can repeat work, waste context/quota, or overwrite its own in-progress changes. The initial task prompt is especially exposed because it deliberately waits for the agent to become idle.

Additional live consequence — OpenCode permission bypass (2026-07-28): the OpenCode E2E task 06FTF9Z8RP5FM4F9M2SKAQNFZ4 was leased to workpc-opencode in /tmp/test-e2e-worktrees/06FTF9Z8RP5FM4F9M2SKAQNFZ4. An operator directly observed OpenCode stop at the interactive permission dialog for its initial git log --oneline -10 && git status --short command, then advance and run that command without operator input. Workpc herdr's live log records Orchestra's agent.prompt calls at 07:37:09 (request IDs 6 and 7), 07:37:10 (ID 8), and 07:37:20 (ID 9); the last two only ended when Orchestra's 10-second RPC deadline disconnected the client. Herdr labels each as changes_ui=true. The active OpenCode manifest contains only permission-detection rules and no auto-approval rule. The task later blocked at the next prompt (~/.claude/RTK.md) when no further prompt retry advanced it.

This does not prove which internal herdr action produced the equivalent of Enter, but it proves the safety boundary is broken: an Orchestra-originated UI-changing agent.prompt operation can advance an OpenCode permission dialog. Treat this as an authorization bypass, not merely duplicate work.

Required fix: make the RPC deadline at least the requested herdr wait (plus a small transport margin), and do not blindly resend agent.prompt after a timeout or other ambiguous post-write error. Record an observable failure instead, or use an idempotency/acknowledgement mechanism if herdr adds one. In addition, never issue agent.prompt while agent_status is blocked (or a pane read matches a permission dialog); treat the condition as requiring explicit operator approval. Do not run further unattended OpenCode E2E tasks until both controls are verified live.

Fix (2026-07-28), in Client.Prompt (internal/herdr/herdr.go:280):

1. The RPC deadline is now derived from the requested wait, not the client default: when the caller's context has no deadline within wait + 5s, Prompt installs one (herdr.go:307-314). A 60s wait.until=idle no longer expires against a 10s transport deadline.
2. Retry is split by error class. A JSON-RPC protocol error is herdr's explicit rejection before it acted and stays retryable for the bounded boot window; a transport failure after the write is ambiguous and is never resent (herdr.go:315-325). Covered by TestPromptDoesNotRetryAmbiguousDelivery, which closes the connection mid-agent.prompt and asserts exactly one agent.prompt is issued.
3. The authorization-bypass half is addressed by inspecting the pane before prompting: Prompt refuses a pane whose agent_status is blocked (herdr.go:285) and refuses any pane whose transcript matches permissionPrompt (herdr.go:292). Approval is now an explicit, separately authorized operation — see the command-channel section below.

Not re-verified live. The original live reproduction was cross-machine, and the "do not run further unattended OpenCode E2E tasks until both controls are verified live" instruction above still stands.

B15 — a coordinator-side block cannot reconcile a later live completion (found live 2026-07-28) — closed (code fix; operator-mediated, not re-verified live)

The same OpenCode E2E task 06FTF9Z8RP5FM4F9M2SKAQNFZ4 was marked TaskBlocked when the coordinator's agent.prompt RPC timed out, despite the agent having received the task. The operator later explicitly approved the one pending ~/.claude/RTK.md read through herdr; OpenCode completed the verification and became idle in pane wN:p1. Orchestra still showed state: "blocked" because Coordinator.Start returned through block() and never persisted a session; its monitor therefore has no live pane to observe or reconcile.

Consequence: the event log can permanently claim a task is blocked when the leased harness has actually completed it. This is not merely stale UI: blocked tasks have no normal completion hook path and their receipt/quota record is lost unless an operator manually corrects the lifecycle.

Required fix: retain enough launch/session identity after an ambiguous lease failure to reconcile agent.get/completion evidence, or introduce a versioned TaskCorrected terminal-state workflow that records the original blocker and evidence without pretending it never happened. A live pane must not be left outside the coordinator's session map simply because prompt delivery timed out.

Fix (2026-07-28), in Coordinator.Start (internal/orchestrator/orchestrator.go:911-922): when the lease call fails but a pane was actually created, the session is completed with HerdrID, TaskFileSHA and ConventionsHash and persisted via rememberSession before block() records TaskBlocked. Reconcile (orchestrator.go:421) explicitly retains mappings for StateBlocked as well as StateLeased, so a restart does not orphan them either. The live pane therefore stays observable through Capture and the monitor.

Scope limit — reconciliation is operator-mediated, not automatic. Nothing promotes a blocked task to completed on its own; the router still never retries a blocked task. What exists now is the evidence and the surface: the retained session plus the complete / release / handoff actions on the UI task endpoint (internal/ui/ui.go:345-361), and the pre-existing versioned TaskCorrected workflow (S8) for recording the correction without pretending the block never happened. That satisfies the "must not be left outside the session map" requirement; it does not implement autonomous reconciliation, which remains unbuilt by choice.

B16 — hardcoded harness name makes each harness globally single-instance (found live 2026-07-28) — closed (code fix; contract test only partially satisfied)

While starting the real OpenCode healthcheck task 06FTFDW22833F1CCB8K43Z8808, Orchestra created worktree/pane wP:p1 but herdr never attached an agent. A direct retry exposed the actual cause: agent.start was sent with name: "opencode", which herdr rejected with agent_name_taken because the completed prior E2E agent in wN:p1 already owned that name. This is not a herdr limit of one OpenCode process per machine: retrying the same kind: "opencode" in wP:p1 with the valid, unique lowercase name oc-06ftfdw22833f1cc immediately started a second attached, idle OpenCode session while wN:p1 remained intact.

Root cause: Client.StartAgent passes the harness kind as both kind and name; every OpenCode lease therefore competes for the same global agent-name slot. The same defect applies to Claude and Codex.

Required fix: keep kind as the configured harness, but derive name from a validated, bounded task/session identifier (lowercase letters, digits, -, _, maximum 32 characters) and persist it in Session for subsequent lifecycle operations. Add a contract test that starts two same- harness sessions in distinct panes and verifies both attach.

Fix (2026-07-28): agentName(harness, taskID) (internal/herdr/herdr.go:452) derives a bounded, validated name — a per-harness prefix (oc/cl/cx), the lowercased task ID with everything outside [a-z0-9_-] collapsed, capped at 32 characters. Over-long IDs keep a leading stem plus an 8-hex-character SHA-256 suffix, because plain truncation made distinct long task IDs collide in herdr's machine-global namespace. StartAgent passes it as name while kind stays the configured harness (herdr.go:393), persists it as Session.AgentName (herdr.go:145, 445), and binds it for prompt routing (herdr.go:446, adapter.go:144). Pane-scoped operations deliberately keep using PaneID — AgentName is only for agent-targeted calls (adapter.go:531).

Requirement not fully met: the requested contract test does not exist. What is covered is name derivation — TestAgentNameIsBoundedAndValid and TestAgentNameLongIDsDoNotCollide — plus TestPromptDoesNotRetryAmbiguous Delivery asserting the prompt target is the unique name. Nothing yet starts two same-harness sessions in distinct panes and verifies both attach, so the agent_name_taken failure mode itself is untested end-to-end.

B17 — opaque harnesses must not author canonical handoff anchors (found live 2026-07-28) — closed on the worker path; open for Design A

The real OpenCode healthcheck run showed that prompting an opaque agent to write the full continuity.Handoff schema is the wrong ownership boundary. The agent wrote JSON that looked plausible but used anchor.sha instead of anchor.git_sha, omitted required top-level fields by nesting them under knowledge, and could not correctly represent the uncommitted handoff file as a dirty-file hash (a file cannot contain the SHA-256 of its own final contents). Repeated corrective prompts made the handoff less reliable and turned Orchestra's protocol details into agent prompt lore.

Required direction: prompt a harness exactly once for a small semantic handoff/review report (what changed, validation evidence, remaining work, review findings). The worker that owns the checkout must then deterministically collect HEAD, branch, dirty paths and SHA-256 values, construct the canonical handoff artifact, and validate it before publication. Canonical handoff files are worker-owned protocol state and must be excluded from their own dirty-file list. Project-level commands and stable expectations (build, test, file ownership/invariants, etc.) belong in AGENTS.md/CLAUDE.md; only task- specific evidence belongs in the handoff.

This cannot be safely implemented in Design A by homesrv: its coordinator does not own the workpc checkout. Expand the bridge into a real worker-side continuity participant (the deferred Design B direction) so all Git-derived handoff facts are gathered and sealed on the machine that hosts the worktree.

Live evidence (OpenCode healthcheck/review run, 2026-07-28): task 06FTFDW22833F1CCB8K43Z8808 initially hit B13/B16: Orchestra created wP:p1 but its hardcoded name: "opencode" conflicted with the prior E2E session. Starting the same kind: "opencode" with unique name oc-06ftfdw22833f1cc attached successfully without closing the prior pane. The implementation agent created executable scripts/healthcheck.sh, validated bash -n, its normal mode, and --help, then committed 672c123 Add scripts/healthcheck.sh. It first authored a lookalike handoff (anchor.sha, incomplete meta, and all protocol fields under knowledge). A distinct reviewer in a second pane (wP:p2) independently repeated the checks, noticed the nesting error, and rewrote the document. Its claimed schema validation checked only JSON syntax and the presence of keys — it did not use continuity.Decode or the actual typed contract: test remains an array where Handoff.Test is a string, and last_result remains a string where Handoff.LastResult is a structured object. This is a concrete proof that a reviewing harness cannot be trusted to validate canonical protocol state by visual shape alone.

The task is deliberately still TaskBlocked from its original failed automatic attach even though its manually-started implementation/review panes are live, a second live confirmation of B15. The reviewer pane is left open for operator inspection; do not close or release it without approval.

Related live observation — unexpected Claude launch explained: Claude pane wM:p1 did not start spontaneously. At 08:00:41Z the expired lease for task 06FTF8CPH3K3DPN6XQA7G3WRQ8 emitted TaskReleased(reason: "lease_expired"); the router re-leased it at 08:00:42Z, and workpc herdr attached Claude at 08:00:43Z. The renewed attempt then hit B14's prompt timeout and became blocked. This is expected retry behavior, but illustrates why lifecycle retries must be visible to operators.

Status (2026-07-28): satisfied on the worker path, unchanged for Design A. cmd/orchestra-worker now implements the required ownership boundary. The harness is asked once for a bounded semantic report; releaseReady (cmd/orchestra-worker/main.go:180) then runs CLIAdapter.Release and herdr.HeadSHA on the machine that hosts the worktree, so HEAD, branch, dirty paths and SHA-256 values are collected locally and the canonical artifact is validated before publication. The handoff-provenance correction documented at the end of this file supplies the parsing and validation half.

The homesrv coordinator's Design A path is unchanged and still cannot satisfy this: it does not own the workpc checkout, so any Git-derived fact it seals is derived from the wrong filesystem. B17 is therefore only closed for tasks that actually run through a worker. The standing rule holds — homesrv must not rotate or clean up a non-local worktree.

B13 — agent.start silently no-ops under back-to-back leases (found live 2026-07-28) — closed (code fix; not yet re-verified live)

Discovered while live-verifying the B12 fix below. B12 itself is confirmed fixed (see that entry), but re-testing it exposed a second, deeper defect that B12's retry logic cannot address.

Three fresh tasks were leased in quick succession against workpc-claude (06FTAKSJZTB73FZQE3QT7XQ1J0/pane wD:p1, 06FTAMFKVEPGPKR0H86C08ZZNG/pane wE:p1, 06FTANAYZPA55BABR2J4MWQ050/pane wF:p1). All three got a worktree.create success and an agent.start call that returned no error. Checked live via pane.get/pane.list (and independently confirmed by the user cd'ing into each worktree by hand):

• wD:p1 (the first, given time alone to settle): a real claude process did eventually attach — agent: claude, agent_status: idle, revision: 2. This is consistent with B12's model: agent.start acknowledges the request but the actual attach is asynchronous, sometimes taking well over a minute.
• wE:p1 and wF:p1 (started moments after wD:p1, while it was presumably still initializing): still no agent attached after 2+ minutes of polling — agent_status: unknown, no agent field, revision never incremented past its creation value. Empty shells, confirmed both via pane.get and by the user directly cding into the worktree and finding no session running.

So agent.start's "success" is not just delayed for these two — it never happened at all, and agent.start never returned an error to tell Orchestra that. One plausible cause: something about launching a second/third claude CLI process while a prior one on the same host/account is still mid-startup (auth handshake, config lock, etc.) silently drops the later request(s) rather than queuing or erroring them. Not confirmed — would need a controlled single-at-a-time repro to isolate from herdr's own internals, which requires spinning up more real panes and wasn't done this pass (see "three orphaned panes" below).

Consequence: unlike B12 (which at least produced an observable TaskBlocked), this failure mode leaves the task leased indefinitely against a pane that will never produce a session, with no error surfaced anywhere — worse than B12 was, because nothing currently distinguishes "still booting, give it more time" from "silently dead, will never start." A client-side retry on agent.start itself (B12's approach) cannot fix this, since the call that should have started the process already returned success with no error to retry on.

Fix: since agent.start's own return value can't be trusted, StartAgent (internal/herdr/herdr.go) no longer treats its success as the end of the story. After agent.start returns without error, it polls pane.get({pane_id}) (new paneAgentAttached helper) until the pane reports a real attached agent (agent non-empty and agent_status present and not "unknown"), bounded by a new agentAttachWindow (90s, longer than bootRetryWindow's 15s since a legitimate attach was observed live taking "well over a minute"). If the window elapses with no attach, StartAgent now returns an explicit error instead of a false success — this is exactly the "recorded/observable failure over silent continue" pattern this repo's CLAUDE.md calls for, and should surface as TaskBlocked through the same path B12's fix already proved reachable.

Not yet done: re-verified against a live herdr instance (the schema for pane.get's result — agent/agent_status fields — was inferred from prose in this file's own B12/B13 narration of live pane.get output, not re-confirmed by a fresh probe of 192.168.1.105:9245; deploy/herdr-schema.json has no entry for pane.get's result shape). go build/go vet/go test ./... all pass, but no test exercises StartAgent's new polling loop directly (existing internal/herdr tests don't call agent.start/pane.get through the fake TCP listener at all). Next session should confirm the pane.get field names live before trusting this closes B13 operationally, and ideally fire a fresh back-to-back-lease test against workpc once that's confirmed.

Side effect of this investigation — three live orphaned panes on workpc, left untouched on purpose: wD:p1 (has a real but abandoned claude session, task now stuck in blocked from before B12's fix), wE:p1 and wF:p1 (empty shells, no agent ever attached). Not cleaned up — same standing rule as the pre-existing wA:p1 stuck task: don't call pane.close/pane.release_agent against live panes without asking first.

---

B12 — StartAgent races the pane's shell readiness (found live 2026-07-28) — closed (this specific race; see B13 for a second issue it exposed)

Discovered during the first real end-to-end run against a freshly rebuilt/ redeployed binary (the running service had been on stale pre-audit commit 325c684 the whole time).

Created a fresh task (06FTAH4MCRCAJYY7V75Z4V2YGG, project test-e2e, capability opencode). It reached TaskLeased onto workpc-opencode (worktree created for real at /tmp/test-e2e-worktrees/<task-id> on workpc, confirmed via a live pane.list — pane wB:p1, correct cwd), then 445ms later hit TaskBlocked:

lease: herdr protocol error: agent target pane wB:p1 is not an available shell

internal/herdr/herdr.go:199 (StartAgent) calls agent.start on the pane Worktree just recorded, with no wait or retry after worktree.create returns. herdr hands the pane back before its shell has actually finished initializing, and agent.start refuses it as not-yet-a-shell. None of B1–B11 exercised this — every prior fix assumed a session already existed (occupancy, rotation, handoff); nothing in the remediation plan tested the very first Lease call against a real herdr pane end-to-end. Phase 0's probing used params:{}/missing-field tricks, never a real worktree.create → agent.start sequence timed against actual pane boot.

Reproduced a second time, 2026-07-28, same session: a second fresh task (06FTAHW8GZP0S6DCHEQ6PJXWB0) leased onto a brand-new pane (wC:p1) and hit the identical error on the same sub-second timeline. Two for two — a systematic race, not a flake.

Consequence: every fresh lease that hits this race goes straight to TaskBlocked, and the router never retries a blocked task (internal/router/router.go has no reference to StateBlocked at all, confirmed by grep) — so it sits there permanently until a human intervenes via the approval surface. This is now the single blocking defect for proving the rest of the remediation plan against a real live task.

Fix, iterated live across three rounds, 2026-07-28:

1. First pass: bounded-retry agent.start (10×, 500ms apart) specifically on the "not an available shell" string. Redeployed and re-tested against a real fresh lease — this cleared that exact error, but the very next call in the sequence (Lease's post-agent.start Prompt) then hit a different transient error, "agent ... is not an active named agent" — same underlying readiness race, one call later, worded differently.
2. Second pass: added the identical string-matched retry to Prompt for that specific message. Redeployed, re-tested — this time hit a third distinct wording, "agent target ... not found", on the same call.
3. Given three different error strings for the same race across three live attempts, string-matching was abandoned as unwinnable (herdr's wording for "not ready yet" isn't fixed enough to enumerate). Final shape: both StartAgent's agent.start call and Prompt's agent.prompt call now retry on any error, bounded by wall-clock time (15s, bootRetryDelay of 500ms) rather than attempt count or specific wording — there is no other legitimate reason a call against a pane/worktree orchestra itself just created would fail immediately. StartAgent still special-cases "already" as a non-error (idempotent re-lease).

Confirmed this closes the original race: task 06FTAKSJZTB73FZQE3QT7XQ1J0 (pane wD:p1) leased successfully, and polling pane.get live shows a real agent: claude / agent_status: idle attach — independently confirmed by the user cd-ing into that worktree by hand. go build/go vet/go test ./... all pass throughout.

Not fully clean, however: re-testing this fix by firing two more tasks shortly after the first one surfaced a second, separate defect — agent.start returning success while never actually starting an agent, with no error for the retry logic to even see. That's B13, above — this entry only covers the specific shell/agent-readiness race B12 was originally filed for, which is fixed; B13 is a distinct root cause the same investigation exposed.

---

Web UI and the capture / approval command channel (uncommitted, 2026-07-28)

Roughly 520 lines of Go across eight tracked files, plus the untracked internal/ui, internal/webui and web/ trees, were added after the last commit and were previously undocumented here. Recording the design so the next session does not have to re-derive it.

Capture/command model (internal/federation/federation.go). Two new worker-scoped maps on Registry. PutCapture stores the latest pane text per (worker, task) and bumps a monotonic Revision only when the text or pane actually changes. Queue accepts only grant_approval / deny_approval, and requires a non-zero CaptureRevision — a command is bound to the exact capture an operator was looking at. CompleteCommand refuses to resolve a command twice.

Worker side (cmd/orchestra-worker/main.go). publishCaptures pushes recent pane text for every held session; runCommands polls /v1/federation/commands and, for each command, re-reads the pane, re-publishes it, and refuses with stale if the resulting revision differs from the one the command was issued against. approvalResponse decides the keystroke: it sends y/n only for a visible [y/n]/(y/n) prompt, and sends bare Enter only for OpenCode's fully-labelled Allow once / Allow always / Reject … enter confirm selector — where Enter is a bounded one-time grant. It refuses to deny through that selector, because doing so would require unobservable navigation. Unknown dialog layouts are rejected outright rather than guessed at. Both branches are covered by TestApprovalResponseOpenCodeAllowOnce and TestWorkerApprovalCommandIsRevisionBoundAndAcknowledged.

This is the right shape for B14's authorization half: approval becomes an explicit, revision-bound, separately-issued operation instead of a side effect of prompting, and it executes on the worker that owns the pane rather than over a coordinator-driven remote socket.

Coordinator side. Coordinator.RespondApproval and herdr.ApprovalResponder (internal/herdr/adapter.go:602) are the local-herdr equivalent, applying the same re-read-before-send rule via an expectedCapture comparison and the same y/n-only refusal. Coordinator.RequestHandoff exposes the handoff request as an operator action without releasing the pane.

Known weaknesses of this channel. The defects found in it are filed individually below — B19 (federated approvals emit no event), B20 (the local capture revision is a timestamp, not a change counter) and B21 (Registry.commands never prunes). Two more, not separately filed:

• Captures and commands are in-memory only. Registry.captures and Registry.commands have no persistence, so a coordinator restart drops pending approvals silently and resets revision counters — which, with a reset counter, could let a stale command match a new capture. This is the other half of B21: the lifecycle forgets across restarts and never forgets within one.
• RespondApproval (coordinator path) has no test, unlike its worker counterpart — and per B20 it is the path whose revision is meaningless, so its text-comparison guard is the only thing actually binding the decision to what the operator saw. That guard is untested.
• web/ is a full Vite/React app whose build output is embedded via internal/webui's go:embed assets/*. The embedded assets are checked in as build output; there is no documented step tying web/ sources to a rebuild of internal/webui/assets, so the two can silently diverge.

Found while implementing B18 (2026-07-28) — all fixed the same day

These were noticed while wiring the session gate, and all five were fixed in the following pass. Each section states the original defect first, then how it was closed.

B19 — federated approvals leave no audit trail — closed (2026-07-28)

Server.action's grant_approval / deny_approval branch (internal/ui/ui.go:301-341) has two exits that are not symmetric. The local-coordinator path calls RespondApproval and then appends ApprovalGranted / ApprovalDenied to the store. The federated path calls s.Workers.Queue(...), writes the command to the response, and returns before appending any event.

So the single most safety-critical operation in the system — injecting a keystroke that advances a permission dialog on another machine — is invisible in the event log precisely when it crosses a machine boundary. There is no record of who approved what, or that an approval happened at all. §7.1's authorization model assumes the event log is the record.

Fixing this needs a decision the local path did not have to make: the command is queued, not executed, so the honest event is a request at queue time plus a resolution when the worker acknowledges (CompleteCommand already carries acknowledged / rejected / stale). Emitting ApprovalGranted at queue time would claim an outcome that has not happened yet.

Fix. The resolution half already existed and this entry understated it: /v1/federation/commands/ in cmd/orchestra/main.go appends ApprovalGranted/ApprovalDenied — but only on status == "acknowledged", so a queued click never claims an outcome. What was missing was the request half. Server.action now appends ApprovalRequested at queue time, with subject_ref set to the command ID that the later resolution event carries, plus decision, worker, pane_id and capture_revision. If that append fails the queued command is immediately resolved rejected, so the worker can never execute a keystroke that left no audit trail. rejected/stale outcomes still emit no resolution event; the ApprovalRequested stays unresolved and surfaces in operations.NeedsAttention, which is the honest reading. Covered by TestFederatedApprovalIsQueuedAtTheWorkerRevisionAndAudited.

B20 — the local capture revision is a timestamp, not a change counter — closed (2026-07-28)

Server.capture (ui.go:80) fabricates Revision: uint64(time.Now().UnixNano()) for the coordinator path. The federation path's revision is a real per-change counter (Registry.PutCapture bumps it only when pane text or pane ID actually changes), and the worker's staleness check depends on that meaning.

Two consequences:

1. The local revision changes on every read, so it conveys nothing about whether the pane changed. The local path is still safe, but only because RespondApproval re-reads and compares capture text — the revision it reports to the browser is decorative.
2. If a task has both a live coordinator session and a published worker capture, capture() prefers the coordinator (ui.go:77) and returns a timestamp revision, which action() then passes to Queue as CaptureRevision. That value can never equal the worker's counter, so the worker resolves every such command stale and the approval silently never happens. The precedence between the two capture sources needs to be explicit rather than incidental.

B21 — Registry.commands is append-only — closed (2026-07-28)

r.commands[worker] is only ever appended to (federation.go:119); CompleteCommand flips a status in place and nothing is ever deleted. Resolved commands accumulate for the process lifetime, and Commands() rescans the entire history on every worker poll. Combined with the in-memory-only storage already noted, the lifecycle is wrong at both ends: it forgets across restarts and never forgets within one.

Fix (within-process half only). Registry.pruneCommands drops resolved commands older than CommandRetention (30 minutes), running on both Queue and Commands, so the per-worker list is bounded and the poll path no longer rescans unbounded history. Pending commands are never pruned, at any age — dropping one would silently discard an operator decision. Covered by TestResolvedCommandsArePrunedButPendingOnesSurvive. The persistence half of this defect is still open: captures and commands remain in-memory only, so a coordinator restart still drops pending approvals silently.

S12 — ORCHESTRA_NTFY_TOKEN serves two unrelated purposes — closed (2026-07-28)

The same env var is the ntfy server credential (main.go:1153, delivery.Ntfy{Token: ...}) and the authz surface credential for authz.Ntfy (main.go:1165). These are unrelated secrets with different trust boundaries: one is handed to a third-party notification server, the other authenticates callers to Orchestra. Setting the former silently makes it a valid inbound credential. They need separate variables.

Fix. The authz.Ntfy surface token is now ORCHESTRA_NTFY_SURFACE_TOKEN; ORCHESTRA_NTFY_TOKEN is once again only the credential handed to the ntfy server. Documented in deploy/orchestra.env.example. Operator action: any deployment that relied on the old dual use must now set the new variable explicitly, or the ntfy surface reverts to having no inbound gate.

S13 — dead auth() middleware in cmd/orchestra/main.go — closed (2026-07-28)

func auth(next http.Handler) (main.go:1212) is defined and never referenced; Go does not flag unused functions, so go vet stays quiet. It reimplements the notify-only rule that authz.HTTP already enforces. It is harmless today and a trap tomorrow — a second, divergent copy of the authorization policy sitting next to the real one. Delete it.

Positive note — B8 is more strongly closed than it was

authz.HTTP downgrades a caller-declared system surface to Web before the token comparison. Previously that landed on a surface whose token was unset in production, i.e. no gate at all. With ORCHESTRA_WEB_TOKEN now mandatory, the downgrade lands on a genuinely gated surface, so the B8 bypass is closed by construction rather than by the header check alone.

---

What's next

In dependency order, not importance order:

1. Finish applying the B18 env change on the deployed instance. The repo-side .orchestra-config/orchestra.env now sets a generated ORCHESTRA_WEB_TOKEN and ORCHESTRA_UI_INSECURE_COOKIE=1, but the unit reads /etc/orchestra/orchestra.env, which is orchestra:orchestra 0600 and unreadable from this sandbox — whether it carries the token could not be confirmed. It was last modified 2026-07-28 14:30 and the service did start at 16:17 with the B18 build absent, so this is unverified either way. Also newly required: ORCHESTRA_NTFY_SURFACE_TOKEN (S12) if the ntfy surface should stay gated, and the token for every existing unauthenticated /v1/ client, which now defaults to the Web surface.
2. Install the new binary and restart. orchestra.service has been stopped since 2026-07-28 19:46 and /usr/local/bin/orchestra predates every fix in this section. Needs privileges the sandbox lacks: go build -o /tmp/orchestra ./cmd/orchestra && sudo install /tmp/orchestra /usr/local/bin/orchestra && sudo systemctl restart orchestra.service.
3. The live re-verification that B13–B17 all still lack, plus a first live exercise of B18–B21. Each of B13–B17 was found live and closed on paper; the code fixes are unproven against a real cross-machine run, and that remains the single biggest gap between this document and reality. Note that no herdr is currently reachable at all — the 16:17 startup logs connection refused for all six, including workpc's, which was live on 2026-07-27 — so a live run needs a herdr brought up first.
4. Two operational faults visible in the journal, unrelated to this audit's defects but blocking a clean live run: every herdr is refusing connections (above), and ntfy delivery is failing 403 Forbidden on every send (16:43 and 16:46), i.e. the ntfy server credential is wrong or expired.
5. The persistence half of B21 — captures and commands are in-memory only, so a restart still silently drops pending approvals. The retention fix bounds growth within a process; it does not make the lifecycle durable.
6. RespondApproval (coordinator path) still has no test, unlike its worker counterpart. With B20 fixed its revision is now meaningful, but its text-comparison guard remains the thing actually binding a decision to what the operator saw, and that guard is untested.

Deliberately not next: building further on Design A's cross-machine calls, and closing B17 for the coordinator path. Both wait on the federation-fork decision below.

---

Closed defects

Each entry is the flattened final state — design + what's verified — not the session-by-session narration. Grouped by original defect ID.

B1 — Occupancy measured from the wrong value (§5.2.1) — closed

CLIAdapter.Occupancy was calling a.Usage(s.PaneID) (a herdr pane id) against readers (ClaudeUsage/CodexUsage/OpenCodeUsage) that want a filesystem path to session state — every call failed and was swallowed by a bare continue. Fixed: herdr.Session.SessionFile is resolved at lease time per harness (Claude: newest-mtime transcript under Claude Code's own project dir; Codex: existing CodexActiveUsage sqlite discovery; opencode: refuses loudly rather than guess, since it needs a live session id not resolvable from the worktree alone). A missing/unreadable session file is now a hard error surfaced via SessionHealth.Occupancy/OccupancyError on GET /v1/tasks/{id}/health, never a silent zero.

Still open: live verification against a real Claude Code session at a known context fill — not possible from this sandbox, blocked further by B13 now that fresh leases don't reliably survive.

B2 — Adapter lookup keyed wrong in 3 of 4 call sites (§5.3, §5.4) — closed

AdapterFactory.Herdrs is keyed by herdr instance id (homesrv-claude); Reconcile/expire/rotate were looking it up by session.Harness (the harness kind, claude) and silently no-op'ing (continue) on every miss — orphaned panes never killed, expired leases never killed their pane, rotation never began. Fixed with a single Coordinator.adapterFor(session) routed through at all four call sites. Regression test registers an adapter under a herdr-id key distinct from the harness kind and asserts rotation still fires.

B3 — Nothing emitted TaskCompleted (§4, §5.2) — closed

Only producer used to be a human calling POST /v1/tasks/{id}/complete. Added POST /v1/harness/complete: a Claude Code Stop hook (deploy/hooks/orchestra-stop.sh) fires every turn boundary but only reports completion once the agent has written a .orchestra-report.md marker at the worktree root (an ordinary turn boundary is a no-op). The server builds the receipt itself from herdr.ClaudeUsage against the transcript (never trusts a self-reported number) and uploads the report body to CAS for report_ref. Gated by optional ORCHESTRA_HARNESS_TOKEN; event appended with Surface: system set directly in Go, consistent with B8. Extended to Codex/opencode via an optional harness field in the request body ("codex" → CodexUsage, "opencode" → OpenCodeUsage).

Still open: no automated test for the HTTP handler (cmd/orchestra/main.go has zero handler test coverage of any kind, pre-existing gap — this follows the existing pattern rather than introducing a one-off harness).

B4 — Router counted rotation as a retry (§5.3, §5.4) — closed

Every TaskReleased (including rotation, which carries a valid handoff_ref) advanced MaxAttempts, and lease also double-counted — a task healthy enough to rotate twice was killed. Now only a release without a handoff_ref (expiry/crash) advances the counter.

B5 — Herdr protocol methods unverified/invented — closed

Confirmed live against 192.168.1.105:9245 (raw JSON-RPC probing, no local herdr CLI available — schema reconstructed from Rust serde's "unknown variant"/"missing field" errors). pane.release, pane.kill, pane.rotation_signal, pane.status do not exist. Real replacements: pane.close({pane_id}) for kill (drop-in); pane.release_agent({pane_id, source, agent}) for release — structurally different, does not return a handoff_ref (herdr never writes handoffs; the agent does, §6.1). RotationSignal interface/method/call site deleted outright (no replacement exists; herdr has no concept of Orchestra rotation).

CLIAdapter.Release now: reads the agent-authored .orchestra-handoff.json, validates with continuity.Decode, cross-checks its anchor SHA against the worktree's real HeadSHA, re-verifies every Anchor.Dirty file hash, scratch-commits any dirty files (continuity.ScratchCommit, idempotent across multiple rotations) and rewrites the handoff's anchor to the new scratch commit before uploading to CAS via continuity.Save (minting the real handoff_ref), and only then calls pane.release_agent — sequenced last so a herdr-side error can't strand an uploaded handoff. Any failure (missing file, invalid schema, anchor mismatch, stale dirty hash, herdr error) is a refusal, which rotate already treats as "retry next tick." CLIAdapter.Lease's bootstrap prompt also fixed to pass time.Minute (not wait=0) for inline-wait, matching Bootstrap and closing the send-into-a-half-rendered-prompt race §5.1 requires guarding against.

Covered by internal/herdr/adapter_test.go against a real git worktree and a fake in-process herdr TCP listener: upload-and-release on a valid handoff, refusal with no handoff file, refusal on anchor mismatch, refusal on stale dirty-file hash, refusal with no CAS configured.

Protocol version confirmed live as a bare JSON number (17), not the string config.jsonc declares — CheckProtocol's raw-bytes fallback happens to compare correctly today; don't "clean up" that code without re-checking this, or it may start doing a real numeric-vs-string comparison and break.

Still open: nothing makes the agent actually write .orchestra-handoff.json unprompted (closed separately, see B6/Phase 4 item 2 below) — that's the piece this entry originally deferred.

B6 — Layer 3 (continuity) was entirely dead code (§6) — closed

Nothing wrote TASK.md, so pickup validation had nothing to check and never ran. Fixed in full:

• GitWorktrees.Create now writes and commits an immutable TASK.md (continuity.RenderTaskFile) into every fresh worktree; continuity.TaskFileHash reads it back for w.TaskFileSHA.
• Coordinator.Start runs continuity.ValidatePickup (anchor SHA + dirty-file hashes + TASK.md hash) before bootstrapping a successor onto a handoff_ref — failure kills the session and emits TaskBlocked instead of trusting an unvalidated ref.
• ScratchCommit wired into Release (see B5) so pickup collapses to a single HEAD compare instead of per-file rehashing.
• CLIAdapter.Bootstrap's prompt rewritten to point the agent at git log/the scratch branch (trust already established by the plane's own ValidatePickup) instead of vague "read the handoff" prose, and deliberately does not claim a GET /v1/artifacts/<ref> endpoint since none exists (/v1/artifacts is POST-only — an earlier draft invented this and was corrected before landing).
• continuity.MarkdownChanges (zero callers, zero tests despite being listed as implemented in an earlier snapshot) was deleted rather than half-wired. §6.3's actual requirement ("notice to agents whose task is adjacent") was rebuilt independently: continuity.ConventionsHash(root) hashes AGENTS.md/CLAUDE.md/VOCAB.md; Coordinator.checkConventions runs every Monitor tick, recomputes the project base repo's hash for every leased session, and on drift calls the optional herdr.ConventionsNotifier capability (in-pane prompt) once per drift.
• Handoff production, the item that most of B6 hinged on: Coordinator.rotate checks the adapter's optional herdr.HandoffRequester capability; if .orchestra-handoff.json is missing, it prompts the agent once (CLIAdapter.RequestHandoff, naming the exact §6.1 JSON shape and explicitly telling the agent not to fabricate SHA/hashes) and skips Release that tick, retrying every subsequent tick — mirrors the .orchestra-report.md/B3 convention exactly. Session.HandoffRequested avoids re-prompting every tick.

Covered by TestGitWorktreesCommitsTaskFile, TestStartBlocksOnInvalidPickup, TestReleaseScratchCommitsDirtyFilesBeforeUpload, TestReleaseRefusesOnStaleDirtyFile, TestConventionsDriftNotifiesActiveSession, TestRotationRequestsHandoffBeforeReleasing (internal/herdr, internal/orchestrator).

Caveat still open: TASK.md hashing is best-effort/untested for the herdr-hosted (WorktreeCreator) worktree path specifically — no adapter or test exercises that path with a real handoff_ref, so pickup validation there runs with an empty taskFileSHA (anchor + dirty-file hashes still checked). Same cross-host caveat as the federation fork, below.

B7 — Quota projection had no producer (§7.2) — closed (post-hoc only)

POST /v1/harness/complete now appends a QuotaReported event (consumed from the same usage.Numerator() used for the receipt), so the router's 5h/weekly filter and the brief's quota_consumed stop evaluating against a permanent zero.

Design gap recorded, not yet built — a live push producer, since a harness that never completes a task cleanly (e.g. the stuck wA pane, see below) currently under-counts its consumption forever:

• Claude: the statusline stdin JSON already carries real .rate_limits.five_hour.used_percentage / .seven_day.used_percentage — confirmed by reading ~/.claude/statusline.sh on this machine. Prefer this over claude.ai/api/.../usage, which needs browser session cookies, not an API key.
• Codex: every token_count event in the rollout carries a rate_limits object with used_percent/window_minutes/resets_at — confirmed directly from a real rollout file. No codex usage subcommand; the rollout tail (or codex app-server, same events live) is the only transport.
• opencode: no first-class quota surface. opencode stats is historical only. Zen's free-tier daily quota arrives as x-ratelimit-* response headers, not a queryable endpoint — would need an opencode plugin to intercept.
• Structural consequence if this is ever built: sumSince's additive model is wrong for a used-percentage feed (need a second Availability reading the latest report per harness, not summing); static quota_limit_5h/weekly become unnecessary for Claude/Codex once real fractions are reported; QuotaWindowLimits{FiveHour,Weekly} is too narrow for Codex's self-describing window_minutes or opencode's daily Zen window.

B8 — Surface: system unauthenticated bypass (§7.1) — closed

X-Orchestra-Surface: system was reachable from any HTTP request header in both authz.HTTP and main.go's surface closure — since no deployment sets ORCHESTRA_SYSTEM_TOKEN, this was an unauthenticated full-control bypass reachable from any LAN caller. Both call sites now downgrade system to web before doing anything else with it. System is only constructible in-process (router leases/failures, coordinator releases/blocks, standup advisory/apply) via the bus-level authz.AuthorizeEvent check now enforced inside store.Append itself.

S1 — duplicate JSON struct tags — closed

Brief.From/To and GitSync.Branch/Head/Status each shared one JSON tag (Go only honors the first json:"..." tag on a combined field declaration), making go vet ./... fail and the brief's git state unparseable. Fixed; go vet ./... passes clean.

S2 + S3 — brief git state and completion receipts — closed

Brief.Git used to read from ORCHESTRA_DATA (the event-log dir, never a git checkout — always "git unavailable") and completions were counted but never surfaced with proof. operations.Brief.Git is now map[string]GitSync keyed by project ID, built from registry.Project.Repo (falling back to a single "default" entry off ORCHESTRA_REPO), with Ahead/Behind vs upstream added. Brief.Receipts []CompletionReceipt pulls report_ref/ receipt straight out of each TaskCompleted event's existing payload.

S4 — notification goroutine died on first send error — closed

delivery.Fanout.Run used to return on the first sender error, permanently killing notifications for the rest of the process's lifetime. Failed sends now go through an OnError hook instead of aborting the loop. Cursor is persisted (SaveCursor → a delivery-cursor file next to ORCHESTRA_DATA), so a restart resumes from the last delivered event instead of re-notifying the entire log from seq 0.

S5 — colliding event IDs on lease — closed

Store.Lease/ExpireLeases set Event.ID to the task id, so every lease of a task produced colliding event IDs. Now uses domain.NewID().

S6 — ingest dedup returned the wrong event with 200 — closed

Dedup path returned nil (success) without appending; the HTTP handler then returned an unrelated event with 201. Added domain.ErrDuplicate and Store.TaskBySource; POST /v1/tasks now returns the existing task with 200 on a duplicate. Every other Append caller (Gitea poll/webhook, JSONL ingest) now treats ErrDuplicate as expected.

S7 — TaskAmended dropped most fields — closed

Only title was ever applied from an amendment payload; due/ description/inherent_priority were accepted, logged, and silently dropped by the projection. Also added the missing Task.Description field entirely (it didn't exist, so TaskCreated dropped it too). Both TaskCreated and TaskAmended now populate all four fields.

S8 — no compensating-event mechanism (§3.1) — closed

Added TaskCorrected: payload requires corrects (the id of the event it repairs) plus at least one change (state, or the existing amend-style fields). Store.Append rejects a corrects that doesn't name a real prior event on the same task; apply applies the changes and clears Lease like every other terminal-state branch. Covered by a mistaken TaskFailed reverted to queued, an unknown-corrects rejection, and both events surviving snapshot+replay.

S9 — duplicate lease-expiry tickers — closed

Coordinator.Monitor's 30s ticker and main.go's own 1s reclaim ticker both called Store.ExpireLeases independently. Previously filed as "harmless — CAS rejects the loser," but the real effect was worse: the coordinator's expire() is the only place that kills the herdr session/pane for an expired lease, and the 1s ticker running 30x more often almost always won the race, leaving the coordinator's own call with nothing left to expire — silently orphaning herdr panes past TTL. Fixed: main.go's ticker skips ExpireLeases entirely when a coordinator is configured (deferring reclaim to the coordinator's loop), keeping only AssignPending as periodic retry.

S10 — federation worker registration had no admission control — closed

Any caller could self-declare a worker id and self-choose a token, and a second caller could silently re-register an existing worker id with a different token, hijacking its identity/capacity. Added Registry.AdmitToken (pre-shared secret via ORCHESTRA_FEDERATION_ADMIT_TOKEN, checked against the registration request's bearer header); same-ID re-registration now requires presenting the existing worker's own token (legitimate restarts still succeed; different-token same-id is rejected as a hijack).

S11 — soft threshold, milestone, thrash, agent-initiated ROTATE (§5.3) — closed

All four of §5.3's rotation triggers now land:

• Soft threshold: Coordinator.Soft (default 0.55, ORCHESTRA_OCCUPANCY_SOFT) — both rotate() and TurnDecision request a handoff once occupancy crosses Soft, before Hard forces one; TurnDecision returns prepare_handoff advisory (task stays leased, no turn boundary required).
• Agent-initiated ROTATE: handoffReason(worktree) reads a written .orchestra-handoff.json; if reason == "manual", both rotate() and TurnDecision skip occupancy and the turn-boundary probe entirely and release immediately — the agent's own handoff is the boundary signal. Shared release-and-certify tail extracted into Coordinator.finishRelease so this path gets the same anchor-safety guarantee as the threshold path.
• Milestone + thrash: needed a transcript/tool-call introspection source this repo didn't have — built new, internal/herdr/activity.go:
  • ToolCall{Name, Kind, Key, Success, IsTest} normalizes one tool/function call across harnesses.
  • ClaudeActivity pairs tool_use/tool_result blocks by tool_use_id from the transcript already opened for occupancy (an unresolved tool_use — mid-turn — is dropped, not reported).
  • CodexActivity — verified against a real live rollout, 2026-07-28, after the original guessed function_call/function_call_output shape turned out not to exist anywhere in ~/.codex/sessions. Real shape: file edits arrive as event_msg/patch_apply_end (has changes + top-level success directly, no pairing needed); shell commands arrive as a freeform response_item/custom_tool_call named "exec" whose input is a JS snippet — codexExecCommand regex-extracts the first embedded cmd:"..."; failure is signaled by the literal output prefix "Script error:" (confirmed for both a JS syntax error and an apply_patch verification failure on this machine's real transcripts). Manually re-ran the rewritten parser against a real multi-hundred-line rollout end-to-end and spot-checked the output by eye, in addition to TestCodexActivityParsesRealRolloutShape / TestCodexActivityMarksScriptErrorAsFailure.
  • OpenCodeActivity refuses outright — opencode's on-disk message storage is only confirmed to carry aggregate token counts, not per-tool-call records; fabricating a parser against an unconfirmed shape would repeat the exact mistake this audit exists to catch.
  • DetectThrash(calls, ThrashConfig) implements all three rules: N consecutive failed test runs (regex-matched build/test commands), the same file edited M times (edit tools only, explicitly excluding Read — caught by an initially-failing test), and an identical tool call repeated K times back-to-back (excluding test re-runs and file reads — also caught by an initially-failing test). Defaults 3/5/4, overridable via Coordinator.Thrash.
  • DetectMilestone(calls) — deliberately narrow: only "the most recent call was a successful git commit." Fuzzier definitions (passing test suite, finished subtask) were not guessed at.
  • CLIAdapter.Activity resolves the session file the same way Occupancy does; ActivityReader is an optional Face-B capability like the others.
  • New ReasonedHandoffRequester/CLIAdapter.RequestHandoffReason names why (thrash's dead ends, milestone reasoning) instead of the generic threshold framing.
  • rotate()/TurnDecision generalize "reason=manual bypasses occupancy" to manual/milestone/thrash alike: checkActivityTriggers runs before falling through to occupancy/soft/hard (thrash takes priority over milestone), and a hit calls requestReasonedHandoff — request only, never release, same as the soft-threshold path.

Covered by internal/herdr/activity_test.go and internal/orchestrator/rotation_test.go's TestActivityTriggersRequestReasonedHandoffWithoutReleasing, TestTurnDecision subtests for soft-threshold and manual-reason, and TestRotationRequestsHandoffBeforeReleasing.

Known remaining limitation: DetectMilestone's single-command extraction only sees the first cmd:"..." in a chained Codex exec script, so a git commit issued later in the same script isn't recognized as the "last call" even though it executed last. Opencode still has no verified per-tool-call source at all.

Also closed this pass (not separately numbered in the original audit)

• Codex/opencode Stop-hook-equivalent scripts. Neither harness has a native Stop hook, so the server-side /v1/harness/turn and /v1/harness/complete (harness dispatch) existed with no caller for either. Added deploy/hooks/orchestra-codex-poll.sh and orchestra-opencode-poll.sh — background poll loops (default 60s, ORCHESTRA_POLL_INTERVAL) that find the newest session-state file (codex: newest rollout-*.jsonl; opencode: newest file under ~/.local/share/opencode/storage/message/), POST to /complete when .orchestra-report.md exists, otherwise POST to /turn and log (not act on) refuse/rotate_now — advisory-only, since there's no real turn boundary to refuse at from outside either harness process without deeper app-server/SSE integration.
• Bus-level authorization. authz.AuthorizeEvent enforced inside store.Append itself (the single choke point every event passes through), not just at HTTP handlers. Event schema bumped to v2, requiring every event to declare a Surface; schema v1 events on disk still replay (tolerant reader).
• Dual quota windows. router.QuotaAvailability tracks a 5-hour rolling window and a 7-day weekly window independently per harness, applying the conservative 80% rule to each separately.
• Turn-boundary detection made observable. An adapter that fails to answer TurnBoundary now blocks that tick's release rather than treating the failure as "safe to proceed," and increments MonitorHealth.TurnBoundaryDegraded.
• Cross-machine lease correctness has a primitive-level test. TestCrossMachineLeaseAnchorAndQuotaArePerHost exercises a lease claimed through the federation worker HTTP API, validates the anchor against that worker's own local checkout, and asserts quota is accounted per-host — proof for the primitives that exist today; does not yet run against two real physical machines (see federation fork, below).
• Fuzz coverage (internal/domain/fuzz_test.go, FuzzValidatePayload/FuzzValidateEvent) for all event types including malformed nested payloads — no panic, always a typed error.
• Multi-repo Gitea ingestion. provider.Gitea gained a Project field and namespaced SourceName(); provider.MultiGitea dispatches by task.Source; LoadGiteaConfigs loads a JSON array of per-project sources. Legacy single-repo env vars still work unchanged. Previously zero tests existed for the Gitea provider at all (an earlier progress.md claim of coverage was inaccurate) — internal/provider/gitea_test.go added.
• Per-project repos. registry.Project gained optional repo/ worktree_root; main.go builds a PerProjectGitWorktrees, falling back to the global default for any project that omits these fields.
• Rotation's TaskReleased payload validity. Coordinator.rotate used to build {"handoff_ref","reason"}, omitting the anchor_sha the spec and domain.ValidatePayload require whenever handoff_ref is present — store.Append would reject it, the error was discarded (if c.Store.Append(e) == nil), and rotation silently never happened, no visible failure. Fixed: herdr.HeadSHA(worktree) populates anchor_sha from the real worktree HEAD before appending; if the anchor can't be read, rotation now skips that tick instead of emitting a payload guaranteed to fail validation. The federation worker release endpoint (/v1/federation/workers/{id}/release) had the identical gap and now requires/forwards a 40-hex-char anchor_sha, rejecting with 400 otherwise.

---

The federation fork — decided 2026-07-27, still the deferred half

Two incompatible federation designs coexist in the tree.

Design A — "drive the remote socket" (currently deployed, clients/herdr-bridge.go): homesrv calls worktree.create/agent.start/ etc. directly on workpc's herdr over TCP as if it were local — meaning anchor validation (git rev-parse HEAD) executes on the wrong machine relative to the actual checkout. Two outcomes if the coordinator's session.Worktree path happens to also exist on homesrv (likely, since every project shares the same directory layout): either HeadSHA errors and rotation silently skips forever, or — the dangerous case — it returns homesrv's HEAD for an unrelated checkout, passing validation while certifying a commit the agent never touched. Same class of bug applies to per-host quota accounting and to cleanupCompleted's git worktree remove, which runs on homesrv for a worktree that lives on workpc.

Design B — "workers pull tasks" (/v1/federation/*): fully built server-side (registration, heartbeat/TTL offline detection, event-cursor polling/ack, lease claim), zero clients — no worker binary exists anywhere in this repo. This is what the spec actually describes (§2.1: "everything crossing a machine boundary is git + a validated artifact, never live state over the wire"), but every endpoint is currently unreachable in the real deployment. Latent, never-surfaced defects in this unused half: offline detection only runs inside Snapshot(), called solely from a GET endpoint — nothing ticks it on its own, so OnOffline (the hook that releases leases held by a vanished worker) only fires if a human hits that endpoint; the registry is in-memory with no persistence, so a restart forgets all workers and cursors.

Decision, unchanged: keep Design A through Phase 5 (single-host concerns — occupancy, Face B, rotation, continuity — are provable on homesrv alone with workpc's herdr as just another pane host); commit to Design B in Phase 6. Two guardrails were meant to land immediately so Design A can't corrupt state in the meantime: (1) refuse to rotate a lease held by a non-local herdr rather than validate against the wrong checkout, since the protocol schema can't run rev-parse where the checkout is; (2) same treatment for cleanupCompleted's worktree removal. Status of these two guardrails is unverified in this pass — re-check internal/orchestrator before assuming they landed; they are not confirmed closed above the way B1–B11/S1–S11 are.

If the worker binary is ever abandoned, delete /v1/federation/* and record the deviation — leaving both designs in place unmarked is explicitly not acceptable; the repo would read as though the spec-conformant path is implemented when nothing can reach it.

---

Live deployment facts (as of 2026-07-28)

• Runs as orchestra.service on homesrv via this machine's own systemd. homesrv has no local herdr running (connection refused on 9245) — only workpc's herdr (192.168.1.105:9245) is live and reachable. main.go only logs herdr connection failures at startup, never successes, so "no log line" for a herdr does not mean it's down.
• Stuck task, deliberately left untouched: workspace wA, task 06FT6CKD9Y98AZRX6X8K3QXFZG, opencode harness, pane wA:p1. From Orchestra's own point of view it is no longer "stuck" — it now reads state: "failed" (retries exhausted, MaxAttempts hit before B4's fix landed). But the underlying herdr pane is still live and agent_status: "blocked", confirmed via a direct agent.get probe — the router gave up and moved on, but nothing ever released or killed the actual agent, confirming the orphaned-pane prediction from B2/B5. Left untouched on purpose: user was asked and chose to leave it rather than have it cleaned up mid-audit. Do not call pane.close/pane.release_agent against it without asking again first.
• First live end-to-end task run (2026-07-28, after rebuilding/redeploying — the running service had been on stale pre-audit commit 325c684) is what surfaced B12 above, since fixed and confirmed live; re-verifying it surfaced B13, still open.
• Three more live orphaned panes from this pass, left untouched: wD:p1 (task 06FTAKSJZTB73FZQE3QT7XQ1J0, worktree /tmp/test-e2e-worktrees/06FTAKSJZTB73FZQE3QT7XQ1J0 — has a real attached claude session but the task itself is stuck blocked from before B12's fix landed), wE:p1 and wF:p1 (tasks 06FTAMFKVEPGPKR0H86C08ZZNG / 06FTANAYZPA55BABR2J4MWQ050 — empty shells, no agent ever attached, see B13). Same rule as wA:p1: don't pane.close/pane.release_agent any of these without asking first.

---

Known open gaps (as of 2026-07-28)

Numbered defects are not repeated here — B19, B20, B21, S12 and S13 are in "Found while implementing B18" above, and the ordering across all of them is in "What's next". This list is the unnumbered residue: conditions that are known, accepted, or not actionable as a single fix.

• B18's env change is applied repo-side only. .orchestra-config/ orchestra.env now sets ORCHESTRA_WEB_TOKEN, but the unit loads /etc/orchestra/orchestra.env, which is not readable or writable from this sandbox. See "What's next" item 1.
• The listener still binds all interfaces; loopback-by-default was considered and not taken (B18).
• TestAssignsByAffinityCapabilityAndConcurrency is flaky — diagnosed and fixed 2026-07-28. The flake was in the test, not in assignment. Store.Tasks() ranges a map, so its order is randomized per call, and the assertion indexed two separate Tasks() calls (s.Tasks()[0] ... && s.Tasks()[1] ...); it failed whenever the two orderings disagreed. Instrumenting the failure showed a valid TaskLeased event and a genuinely leased task on every "failing" run. It now snapshots once and asserts the real invariant — concurrency 1 means exactly one of the two tasks is leased — and passes at -count=60. Worth recording that which of two equal-priority tasks wins is still nondeterministic, since sort.SliceStable is applied to a randomly ordered slice; that is a real property of the router, not a test artifact.
• B14/B15/B16 are code-fixed but not re-verified live, and B17 is closed only for tasks that run through a worker. Each was found live, so a code fix plus unit tests is weaker evidence than the failure that produced it. In particular B16 still lacks the requested two-pane same-harness attach contract test, so its actual agent_name_taken failure mode is untested.
• B15's reconciliation is operator-mediated only. Blocked tasks keep their session and are correctable through the UI, but nothing promotes them automatically and the router still never retries a blocked task.
• web/ build output and internal/webui/assets can silently diverge; no documented rebuild step ties them together.
• Cross-machine lease correctness proven at the primitive level only — needs an actual homesrv/workpc pair over the real mesh; this repo cannot exercise that by itself.
• Turn-boundary Face B degrades to occupancy-only for adapters that don't implement it, by design — degradation is observable (MonitorHealth.TurnBoundaryDegraded) and blocks-on-failure, but whether Claude/Codex/opencode's native hooks are wired in a live deployment is a deployment-config fact, not provable from source alone.
• Quota has no live push producer — only the post-hoc /v1/harness/complete producer exists; see B7's "design consequences" for the (unbuilt) Claude statusline / Codex rollout-tail feeds.
• Federation guardrails' landed-status is unverified — see federation fork section above.
• opencode has no verified per-tool-call activity source and no first-class quota surface (Zen headers only, would need a plugin) — both named, not attempted, in S11/B7.
• DetectMilestone's single-command extraction can miss a Codex commit issued later in a chained exec script.

Everything else audited (provider layer, continuity schema/validation, router matching, delivery, federation registration primitives) was spot-checked against the code and its tests and matched described behavior.

---

Handoff provenance correction (2026-07-28)

The previous rotation path had a primacy inversion: the harness wrote a free-form Markdown “semantic report”, while canonicalHandoff fabricated the canonical fields around it (goal, done_when, a circular action and a cat command) and placed the entire report in one remaining element. That made the apparent schema a wrapper around prose rather than an authoritative handoff.

The release path now asks the harness for only six labelled, bounded answers: NEXT, WHY, REMAINING, DEAD ENDS, OPEN Q, and LEARNED, each with a NONE escape. canonicalHandoff parses those answers and derives only worker-owned facts (Git anchor, dirty state, metadata, and last observed harness command). It never invents task intent. goal and done_when were removed from continuity.Handoff; pickup must obtain task scope and success criteria from immutable TASK.md.

continuity.Handoff.Validate is the shared semantic gate for both release and pickup. Strict decoding rejects removed/unknown fields, and validation rejects circular handoff actions, commands that point at report/handoff files, oversize or Markdown-smuggled authored list items, and malformed dead ends. A failed release remains refused and is retried by the existing rotation loop; a bad artifact that somehow reaches CAS is also refused at pickup. The fuzzy “empty dead ends after a non-trivial later rotation” signal is intentionally not yet enforced: no reliable diff-size/rotation-index evidence is available at this validation boundary, so inventing a hard rule would create false refusals.

Verified after the change with go test ./..., go build ./..., go vet ./..., and git diff --check.