Run 9's research seal was refused for writing "F1" as a finding id. The rule
is real and the message was precise, but no brief had ever stated it: the
shape block shows keys and types, and a format constraint is neither.
The agent recovered in fifteen seconds, so this cost one boundary rather than
a run. It is still a refusal nobody had to earn, and the same shape hid F38 a
few runs ago.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
Run 8 rendered three phases as "verified at 14654d6cab63, stale because the
tree is now at 56d9b9acdd59" and then printed "Every phase is verified" under
them. The staleness was honest and the conclusion was not.
Stale still counts as done. The usual reason a phase goes stale is the next
phase's own commit, and treating that as unverified would send the implementer
back to phase 1 after every commit. What the successor needed was the names:
which phases were verified against code that has since changed, and an
instruction to reverify what its work touches.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
The gate had two live consumers and no producer. store.go resolves a pending
sign-off by subject, and manuallySignedOff checks for a prior one, but every
comment Orchestra imported was hardcoded to operator_instruction. A phase
carrying a manual step could reach awaiting_manual_verification and never
leave it.
A comment whose first line reads "orchestra verify <phase-id>" now approves
that phase. Orchestra supplies the plan ref from the task's own accepted plan,
so the approval binds to the plan that was current when the human wrote it.
Everything else still lands under operator_instruction, which is what keeps a
generic "looks good" from satisfying a gate nobody was discussing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
Delivery merges the task branch, and that branch carries the "orchestra:
TASK.md" commit. Master therefore ends up holding the previous task's TASK.md,
and the next worktree branches from it.
writeTaskFile returned early on os.Stat, so it left that inherited file in
place. The worker then hashed the current task and every immutability check
failed with "TASK.md changed" against a hash for a task nobody was running.
Releases failed, rotation never relaunched, phase requests were never read, and
the task died on retry_limit without leaving research.
Presence is not identity. Compare content, and rewrite when it differs.
Found live in run 7, the first task to start after run 6's pull request merged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
An implementer that finds the plan contradicted by the code had two options,
both bad: work around it silently, or improvise a different plan inside the
phase meant to execute one. PlanMismatch is the third.
The report carries an observation and nothing else. It may not propose a
replacement plan, because writing the next plan is the planning phase's work.
requested_action stays advisory: replan, research, or human_decision is a
recommendation, and Orchestra decides.
Staleness is checked before anything is recorded. A report names the plan ref
and the commit it was written against, both filled by the worker from what it
can verify rather than from what the agent asserted. A report against an older
plan says nothing about the current one, and one against an older tree may
already be fixed. Neither is replayed.
The reducer keeps two things apart that are easy to conflate:
mismatch recorded != plan superseded
A plan stops being accepted only when a replacement is actually sealed, so an
abandoned replan leaves the accepted plan and its verified progress intact. On
a real re-seal the old ref moves to PlanHistory and its progress stops counting,
while the verification events stay in the log as provenance.
human_decision never reopens. It blocks with a packet stating what was observed
and what it contradicts, and a human answer can resolve the contradiction
without resealing anything: the plan, its progress and the phase all survive,
and the answer outranks the plan where they differ. Turning every ambiguity
into a replan would put the planner above the person who set the goal.
The backward edge is Orchestra's alone. CanReopenPhase is separate from
CanTransitionPhase, which every path validating an agent's request uses, so
phase-request.json still refuses a move back. An agent asks by reporting a
mismatch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
A detailed plan that nothing enforces is a document. This makes the phases
executable: the implementer may write exactly one status, and every other
status is a conclusion Orchestra reaches by running the plan's own commands.
agent may request: ready_for_verification
agent may not assert: verified, awaiting_manual_verification, failed, skipped
The worker resolves commands from the coordinator, never from the request, so a
request cannot smuggle in a command the planner did not write. They run as argv
through exec with Dir set to the worktree, which is the quality gate's existing
envelope and not a weaker one. There is no shell, so a pipe is a literal
argument.
Project policy decides executable reach. registry.Project.Verification matches
argv positionally, and an absent policy refuses everything: a plan command is
agent-authored, so inheriting the operator-authored gate's reach by default
would be the wrong direction to fail in. A refused command is refused before
anything runs, and the refusal names the project and the command so the planner
learns its real reach.
Two bindings make the record mean something later. PlanRef, so progress earned
under plan A cannot survive into plan B. AtSHA, so "verified" does not outlive
the code that made it true: a record whose commit has moved is retained as
provenance and rendered as stale, never as a claim about the current tree.
Both are the same failure this codebase already fixed for reviews, which bind
to the commit they examined.
Manual steps hold a phase at awaiting_manual_verification. The sign-off is an
ordinary human decision whose subject carries the plan ref and the phase id, so
a later "looks good" on an unrelated thread cannot satisfy a gate nobody was
discussing.
A plan sealed before plan.md declares no executable unit, and says so: the
implement context states that phase progress is unavailable and the work
continues under the old semantics. Inventing phases it never had would be worse
than admitting it has none.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
The plan artifact was Changes{Target,Intent} plus three string lists, every
entry capped at 500 single-line characters. That bound makes a specification
impossible: a phase cannot carry a code block, a paragraph of reasoning, or a
verification command with its own argument list. renderSealed then flattened
what little survived through collapse(), so an implement session received a
summary of a summary.
plan.md replaces it. Markdown, 128 KiB, no per-line cap, sealed through the
existing path under the existing PlanRef. The parser enforces the structure the
brief states: required sections, phases numbered from 1 with no gaps, Files,
Changes and Verification per phase, and at least one automated or manual check,
because a phase nobody can verify can never be established as done. Automated
entries are JSON argv arrays, so a pipe is a literal argument rather than an
operator. Headings inside fenced blocks are content, so a plan may show
markdown without parsing its own example.
Citations resolve at seal time against the accepted research, on the
coordinator, which is the only party holding ResearchRef. A plan resting on a
finding nobody recorded fails on the planner while its session is still alive
to be told.
The plan now renders byte for byte into the implement launch, and a rotated
successor receives the same complete document. That is the property the whole
change exists for. collapse() stays for research findings, which really are
short claims.
DecodeStoredPlan reads pre-markdown refs and renders them into the same type,
labelled, so nothing downstream branches on which era a plan came from. A
legacy plan carries no phases, which is honest: the old artifact never named an
executable unit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
The brief at agentctx.go:167 advertised findings[].id and findings[].confidence
to every research session. The struct carried neither, so encoding/json dropped
both on every seal, silently, for as long as the schema has existed. A plan
phase had nothing stable to cite and no way to tell an observation from an
assumption.
Finding gains ID and Confidence. Ids are unique within an artifact and shaped
so "research:<id>" is unambiguous in plan prose. Confidence is fact, inference,
or assumption, matching the labels the output style already uses.
DecodeStoredResearch reads what is already in the CAS and backfills both.
Refusing an artifact sealed before this change would block every task whose
research predates it, including at rotation, where the agent that could fix it
is already gone. A backfilled finding is labelled inference rather than fact:
the old schema required evidence and made no verification claim, so upgrading
it on the way in would be the same class of lie this commit removes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
The issues endpoint returns pull requests alongside issues, and nothing
filtered them. The first submission this deployment ever made, kami/test-e2e#8,
came straight back as task 06G4E83E4KRXM8DS90M2648MGM with the submission
packet as its description. That task would have implemented, reviewed and
submitted again, opening a pull request per cycle.
The webhook had the same hole from the other side: a pull_request delivery
leaves the issue key empty, so it would have appended a task numbered 0 with
no title. Both routes now refuse a pull request, and the poll count reports
what was ingested rather than what was listed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
The completion tail ended at TaskCompleted with no pull request. Nothing in
the running system ever called the review or submission endpoints: the whole
event log holds zero ReviewRecorded and zero TaskSubmitted, so the merge
reflection, the publisher and the human trust boundary had no entry point.
Four links, in the order the tail needs them:
- finalize commits first and runs the quality gate against the committed
tree, so the gate result is bound to the commit being submitted.
CheckSubmission requires gate sha, review sha and head sha to be one
commit, which a gate run on the pre-commit tree can never satisfy.
- The worker seals the reviewer's findings and submits, through a new
/v1/federation/workers/<id>/submit. A blocking review returns the task to
implementation instead; a project with no forge still completes directly.
- The reviewing session is told where findings go. The brief asked for
findings and named no file, and it described a diff nobody supplied.
- GiteaPublisher.Push asks the forge what the branch holds before reaching
for a local checkout. A worker-owned worktree is on another machine and
has already pushed the commit; the coordinator has no such directory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
F41, live on run 5: the review agent produced nothing after 02:12 and the
02:46 renewal was granted anyway. The progress digest covered Claude Code's
status footer, and one of its fields ticked inside the window. Reproduced the
worker's stored progress_sha byte for byte from the live pane, so the branch
taken was progress != ProgressSHA, not IsBusy and not an empty baseline.
PaneProgress now cuts from the editor's lower rule and trims the spinner
summary and version notice above it. AgentStatus still reads the raw capture,
so the busy markers living in the footer are unaffected.
Lease TTL moves to 5 minutes, from domain.LeaseTTL, with renewal at half of
it. Reclaiming a stalled pane happens only at expiry, and 30 minutes per
window made run 5's stall unbounded in practice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CVbaKucEYBjMqVeUgJUsc1
The worker finalises a task when .orchestra/done appears. No brief ever named
that file: grepping a rendered launch.md for it returned nothing, in any phase.
review is terminal. Its only legal move is backwards to implement, so a review
that passes has nothing to ask for and, until now, nothing to write either. Run
5 halted exactly there after four clean rotations, with no error anywhere,
because stopping was the correct reading of its instructions.
The brief now names the marker in the terminal phase, says when to write it,
and says it is exclusive with asking to go back. The test asserts the negative
too, so a phase that can still ask is never told to finish instead.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
Two halves of the same failure, live on run 5.
F38: the phase brief named .orchestra/research.json and described its contents
in prose, never its schema. The agent guessed dead_ends as strings where the
decoder wants {tried, why_failed} objects. The brief now carries the shape, and
a test decodes each documented shape with the same function the worker uses, so
a struct change that is not mirrored fails the build.
F39: the local artifact check refused the request through recordError alone.
answerRefusedPhase only ran on a coordinator 409, so a decode failure told the
agent nothing. The session sat at a boundary rewriting nothing, which is the
silent-loop shape the comment above that block warns about, reached by the one
path with no delivery. Both local refusals now reach the agent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
Handoff.Command is the last command observed in the pane. The handoff prompt
tells the agent to write .orchestra-handoff-report.md and stop, so that write
is almost always the last command there. Validate then rejects it as circular:
"must not point to a handoff or report".
Every phase rotation therefore failed on Orchestra's own instruction. Live on
run 5, one stage past F36.
lastObservedCommand now skips commands the validator would call circular, so
the successor gets the last command that was real work. The rule stays in
continuity and is asked, not restated, which is how F36 happened.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
handoffReason has emitted "phase_changed" since phase rotations landed. The
continuity validator's reason list was never extended, so every phase rotation
built a handoff it then refused as "invalid handoff meta".
Live on run 5: with F31 clearing the parse ahead of it, the release reached
this and stopped at phase "prepared" with
"adapter: upload handoff: invalid handoff meta".
The test asserts the property rather than the constant: every reason the
adapter can produce must survive Validate, including its fallback.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
Two gaps F33 left behind.
F34: the renewal gate exempts a lease with no baseline, so a pane that opened
and never started got a full free renewal period. That is the exact case the
gate exists to catch, and it happened live at 01:33:37: the stuck run 5 lease
renewed to 22:03 on a pane that had not moved since 01:13. Baseline the
progress hash at launch, where the pane is already being read.
F35: nothing could re-poke a live pane. Orchestra can put text in an editor and
be wrong about whether it landed, and the only recovery was to destroy the
lease and wait out expiry, roughly an hour. The new "resubmit" action presses
Enter on text Orchestra itself submitted. It changes no lifecycle state so it
appends no event, and it is fenced like an approval: a live worker-owned pane
at the capture revision the operator was looking at.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
ConfirmInput's first poll can run before the TUI renders the pasted text. It
then finds an empty editor, falls through to the idle branch, and reports
confirmation=editor_cleared. The launch text sits unsent for the whole lease
while the coordinator believes the agent is working.
Run 5 died on exactly that: first_submit_at 21:13:19.570, confirmed_at
21:13:19.577. Seven milliseconds. The two launches that worked took ~500ms and
a second Enter, so the difference was scheduling luck.
An empty editor is only proof once it has held the text, or once it has stayed
empty past a settle window. Text seen and then gone still confirms at once, and
so do busy, blocked and queued. Waiting only happens in the never-observed
case, which is the one that cannot be told apart from a slow render.
With this, the failing case reaches the existing resubmit path instead: the
text appears, is recognised as unsubmitted, and Enter is resent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
A running but empty server answers "no current target". hasSession only knew
"can't find session", "no server running" and "no sessions", so it returned
that as a real error. Kill then failed for a pane that was already gone, the
quarantine never cleared, and the worker's only session slot stayed pinned.
This is a side effect of dea56e4. Before the runtime became its own unit the
server exited with its last session and answered "no server running", which
hasSession already handled, so this reply had never been produced. AgentStatus
shares the same helper and now reports "exited" instead of erroring.
Found live: the F28 disposable task's session survived its own block on
93338b7, quarantined and unreapable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
parseHandoffAnswer joins the agent's NEXT and WHY answers with " — ", and the
prompt asks for a sentence each without naming any budget. Validate then held
that join to one authored line's 200 characters. Two ordinary sentences do not
fit, so every rotation failed.
The failure was invisible twice over. The message said "prose smuggled into
list", which named a branch the answer cannot reach: parseHandoffAnswer splits
on newlines and trims, so no authored field ever contains "\n#". The only
reachable cause was length, and the agent was never told what to shorten.
Seen live on two tasks, and it left the release transaction stuck at "prepared"
that pinned workpc-claude's only session slot (F30).
Give Action the budget of both lines, name the length in the error, and put the
limit in the prompt the agent actually reads.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
The store fences TaskReleased, TaskBlocked and TaskCompleted on a leased task
against the live harness_id and lease_epoch. The UI action handler sent
neither, so all three returned 409 on exactly the tasks the UI listed them as
enabled for. Found live: eight block attempts against a stuck run at a stable
version, all 409 "task version conflict".
The fence is there to reject a stale writer, not the operator. Carry the lease
read at the top of the handler. The version CAS on the append still rejects a
racing write.
Also initialise body when the request carries none. The block path wrote
block_reason into a nil map.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
F25. rotationTick returned early for claude before reaching federatedTurn,
which has one call site below that return. On the harness both burn-in runs
used, no phase request could ever be read and every human decision recorded
against a live session went undelivered. Claude still skips the occupancy
state machine below, because it owns its context rollover through the
installed hook. A turn boundary is not a rotation.
F26. The phase brief listed every domain-legal target, so run 4's frame
session read "research, implement" and asked for implement, which the
project's path refuses. The path is Orchestra's to know: the brief now names
one step and says a wrong target comes back with the right one.
F27. A refused request only reached recordError, leaving the agent to rewrite
the same rejected file forever with nothing telling it why. federation.
StatusError makes a 409 classifiable, and the refusal is delivered through
sendPrompt under the F20 guarantee. A transport failure is not an answer: the
request survives and the agent is told nothing.
The F25 regression test fails against the unfixed rotationTick.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
The brief told the agent to ask for a phase change and never carried the
asking. The agent asked in prose, no code represented the request, and the
session idled until its lease expired. That is what failed run 3.
F21. The agent asks with .orchestra/phase-request.json, and seals
research.json or plan.json where the phase it is leaving produces one. At a
verified turn boundary the worker checks the phase belief, the transition and
the artifact, then calls the coordinator with its lease epoch and a derived
operation id. AdvanceWorkPhase is unchanged, so a request cannot reach a move
the operator surface could not also make. Redelivery is idempotent.
F22. A session now records the phase it was launched to run. One that no
longer matches its task rotates with reason phase_changed, whether this worker
asked for the change or an operator made it.
F20. CLIAdapter.prompt sent handoff and rotation prompts without confirming
them, which is the failure F20 exists to catch. Fixed at the shared call site.
F23 needed no change. Issue comments already become decisions with no
submission, through Reconciler.Reconcile at PreLease and at every turn
boundary. The earlier finding searched internal/operations alone and was
wrong. Tests now cover the boundary it turns on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011xsXyr5J1RACo71YeKG3Pu
F20. Only the launch confirmed its submit. A decision notice at a turn
boundary, and /clear or @HANDOFF.md during a context reset, were
fire-and-forget through the same transport that loses an Enter often enough
that the launch needed three resubmits. A lost Enter on the context-reset path
is the worst of them: it strands the session mid-rollover and nothing retries
it. LaunchConfirmer is therefore InputConfirmer, ConfirmLaunch is ConfirmInput,
and sendPrompt and sendLine both go through it.
Orchestra does not try to guarantee delivery of input it did not originate.
But it must never read that input as work, which is the F16 half. Burn-in run
3 stalled with an unexplained "go ahead and implement it" in the editor, and
the renewal check hashed the whole capture, so those keystrokes read as
progress and the lease kept renewing around an idle agent. PaneProgress drops
input lines from the capture, which the -J join makes exact: a wrapped input
block is one line beginning with the prompt marker.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A task that reached the router's MaxAttempts was permanently terminal.
TaskReleased only ever increments Attempt, TaskCorrected could not touch it,
and no HTTP route emitted a correction at all. The only way to work an
exhausted issue again was to invent a second task for it, which defeats
(source, external_id) dedupe and abandons the task's own history.
POST /v1/tasks/{id}/retry, full-control surfaces only. It requires the task
to be failed, unleased, and failed with reason retry_limit: restoring a retry
budget is not an answer to a failure that was not the budget running out. The
effect is one TaskCorrected naming that failure, setting state queued and
attempt 0 and clearing next_retry_at, failure_class and last_error. Task id,
source pair, goal, acceptance, decisions, work phase and artifact refs all
stay, and the original failure events stay in the log.
operation_id is required and makes the call idempotent, so a repeated request
cannot reset an attempt that has since started running.
This is RetryTask, not a generic correction endpoint: arbitrary task mutation
over HTTP is a different and much larger authority. It also does not address
F9, which is an operator releasing a lease someone else owns.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Burn-in run 3 failed three times with prompt_not_submitted: the launch text
reached the editor every attempt and the following Enter never took effect.
Six isolated probes of the same code path, environment and worktree all
submitted on the first Enter, so the submit is not deterministic and waiting
for it to land is not enough.
F17 first. pendingInput scanned every line beginning with the prompt marker,
but a queued or already-accepted message renders with the same prefix. Only
the editor that owns the pane cursor is unsubmitted input, so inputState asks
tmux for the cursor row and reads the editor around it, joining soft-wrapped
rows.
On that footing ConfirmLaunch becomes an active submit protocol: resend Enter
while the live editor still holds exactly what was submitted, at most three
times and no closer together than two poll intervals, then observe until the
deadline. Queued input confirms rather than fails. The evidence records
confirmation kind, submit_attempts and both timestamps, so a harness that
needs a second Enter is distinguishable from one that needs none.
Verified live against a real Claude Code pane: confirmation=editor_cleared
submit_attempts=1, no spurious resend.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Burn-in run 2 recorded TaskLaunchAcknowledged, opened a pane, and ran nothing
for fifteen minutes. The launch instruction sat in Claude Code's input editor
as "[Pasted text #1 +66 lines]" at zero tokens and zero elapsed. Two separate
bugs produced that.
The transport was wrong for the harness. TmuxBackend.Prompt writes the whole
instruction with send-keys -l and then sends Enter, and the TUI coalesces the
fast multi-line write into a paste that absorbs the following Enter. Launch
transport is now a backend property rather than one universal prompt format:
claude on tmux submits a single line pointing at .orchestra/launch.md, every
other harness keeps the inline path it was verified on. agentctx is unchanged
and the file still holds the exact bytes Orchestra rendered, so what the agent
receives is identical either way. Under the file transport a failed write is
now a failed launch, because there the file is the instruction.
The acknowledgement was also wrong. It meant "Prompt returned nil", not "the
harness accepted the prompt". Backends may now implement ConfirmLaunch, and
the tmux one polls until the input editor clears and the agent is observably
busy, blocked on approval, or at least no longer holding the text. An editor
that still holds the prompt at the deadline is a definite failure. The worker
kills the pane, drops the session so the retry starts clean, and returns
ErrPromptNotSubmitted, which classifies as prompt_not_submitted rather than
launch_uncertain. That class already falls through to TaskReleased, so the
existing retry path takes it and no lease is held on a launch that never
happened.
The confirmation bound is tunable because how fast a terminal harness reacts
is a property of the host. It is not a sleep before the submit: the submit is
deterministic, and this waits for the harness to visibly react to it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gitea ingest set title, description and capability but never acceptance, so
every Gitea-sourced task rendered "Acceptance: Not stated." however carefully
its body was written. F2 restored the body; it did not wire the field.
The convention is deliberately tiny. A markdown heading spelled "acceptance"
or "acceptance criteria" opens the section, bullet and checklist items until
the next heading become the criteria in order, checkbox markers are stripped,
empty items are dropped, and the whole section leaves the description so a
criterion is never also read as instruction prose. Nothing else is recognized:
inferring acceptance from arbitrary prose would eventually invent a
requirement, and a fabricated criterion outranks every human decision beneath
it in the authority order.
An absent section yields no acceptance, which is not an ingestion failure. The
task renders "Not stated." and the frame phase resolves it through the
decision-request path, where a human answers instead of the parser guessing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Burn-in run 2 ingested its task and then sat queued forever. Every herdr in
the live config declares quota_limit_5h and quota_limit_weekly, the event log
holds zero QuotaReported events, and QuotaSince reported an empty window as
unknown. QuotaAvailability fails closed on unknown, so no harness could ever
be leased, and the only producer of a receipt is a completed lease.
The event log is Orchestra's whole accounting source, so a window holding no
receipts is observable zero consumption. QuotaSince now reports known for an
empty window and for a harness that has never reported. A receipt that
declares its own consumption unknown still fails closed.
The refusal also lied about its cause. federatedAvailability collapsed a base
gate refusal into the federation health string, so router health said "stale
heartbeat or unhealthy local backend" while the heartbeat was one second old.
Availability gates now name themselves through an optional
ReasonedAvailability contract: quota refusals say whether usage is unknown or
the window is exhausted and by how much, and worker refusals distinguish an
unregistered worker, a never-probed backend, a stale heartbeat, a stale
health check, and an unreachable backend.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The API was in a restart loop, exiting with `invalid event: until_ns required`.
ValidateEvent compared until_ns against time.Now() for TaskLeased and
TaskLeaseRenewed, so a lease event that was valid when written failed validation
once it expired. store.Open replays the log tail after the snapshot and
log.Fatal's on the first invalid event, so the coordinator refused its own
history and could not start.
Validation of a durable event must be time-independent. Well-formedness is this
function's question; freshness belongs to Store.Lease and Store.ExpireLeases,
which compute until_ns themselves.
Latent since the field was introduced. It needed a renewal in the post-snapshot
tail plus a restart after that renewal expired.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A task in retry backoff was filtered out before the candidate loop, so it
recorded no rejection at all: queued, apparently assignable, and silent. That is
the exact shape that made F5 take a live session to diagnose. It now reports
"retry backoff until <time>".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
F7, security. An unset surface token makes the middleware skip its check, so a
full-control surface with no credential is an open control plane rather than a
closed one. With ORCHESTRA_TUI_TOKEN unset, any LAN caller could lease, release,
complete or block any task by declaring one header, which is how this session's
manual leases were issued. authz.RequireCredentials now refuses startup instead
of logging. Web is exempt: Sessions makes its login mandatory.
F5, lifecycle. router.go's silent `continue` was the first bug, not the
predicate behind it. Every eligibility gate now records a router.Rejection with
task, herdr and reason, exposed at GET /v1/router/health, reset per pass. No
gate was weakened: a direct Store.Lease succeeding proves the lease path, not
that eligibility should have selected that worker.
F8, correctness. Reconcile iterated every configured source for every task, so a
task's external id was looked up in whatever repository each source pointed at.
Once two repositories share an issue number, an unrelated human comment becomes
an authoritative decision for the wrong task. Reconciliation is now bound to
task.Source, the provider:project identity the ingest stamped, and a source that
cannot prove it owns the task is skipped. A task with no matching source
reconciles to nothing and still launches, because nothing to import is not a
failure to read.
The integration fixture ingested from "jsonl" while reconciling from "gitea",
which is exactly the shape F8 makes impossible; it now ingests from the source
it reconciles.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The .orchestra directory showed up as untracked work in the very worktree whose
own launch context said "uncommitted changes: false". It would have polluted
the quality gate, the review diff, and the agent's reading of git status. A
.gitignore of "*" inside the directory ignores it including itself. The
worker's done marker lives there too and had the same problem.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Found by reading .orchestra/launch.md on the first burn-in task. The rendered
context said `Acceptance: Not stated.` and carried the issue title as the whole
goal, because Gitea.event parsed the issue body and then dropped it from the
TaskCreated payload.
The body is the task's own statement of what it wants, which every rendered
context ranks above continuity and below only a human decision. The store
already reads `description` from TaskCreated; only the provider was silent.
Every Gitea-sourced task so far has therefore run on its title alone. Classify
as an authority bug, not a model-following failure: no agent could have known
what it was not told.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Found live on the first burn-in task. Every federated launch failed with
`effective intent: federation: 401 Unauthorized: unauthorized surface`, and the
task nacked back to queued.
GET /v1/tasks/<id>/intent was added so a worker renders its own launch
instruction from the reduced intent, but authz.HTTPWithSessions never got the
matching worker-path exemption. An unlabelled request defaults to the Web
surface, which is session-gated, so the one caller the endpoint exists for could
never reach it.
Twenty test packages passed throughout, because the authz tests covered the
surface capability table and not the worker path list. Only the live owner path
established this.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The v3 stack, previously an uncommitted working tree, plus this session's two
units and the burn-in instrument. This commit is the burn-in build identity:
coordinator and worker must both report this revision before a task is created.
Workflow (earlier sessions, uncommitted until now): human decision events and
reduction, source cursors and reconcile-before-launch, turn-boundary
reconciliation, internal/agentctx as the single renderer, ace-fca phases with
sealed artifacts, the trajectory gate, bounded grilling, independent review,
task pr enforcement, and human review reflection.
Capability restrictions at the agent boundary: an authz.Agent surface at
GatedWrite may ask and may not act. It also fixes two bugs the unit exposed --
gated surfaces could not reach the two endpoints written for them, and
RequestHumanDecision would block an unowned task while rejecting a question
from the session that did own it.
Turn-boundary reconcile-failure escalation: a streak of consecutive failures
asks the session to hand off, fenced on the lease epoch, with reconcile_failure
as a real handoff reason. The worker was dropping the coordinator's verdict on
the floor; it now acts on it.
Burn-in: herdr.WriteLaunchContext dumps the exact agentctx.Build result to
<worktree>/.orchestra/launch.md at every launch, local and federated. BURNIN.md
is the runbook. deploy/build.sh stamps both binaries from one commit.
go build, go vet and go test ./... pass, 20 packages.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Acts on the 2026-07-30 senior review (REVIEW.md findings 1, 2, 4, 5, 7).
Docs (finding 1): CLAUDE.md and AGENTS.md both claimed Design B "has zero
clients - no worker binary exists". cmd/orchestra-worker/main.go is the
deployed worker, and the non-local-herdr guardrail has landed in
Coordinator.adapterFor. Both sections rewritten; AUDIT.md gains a matching
federation-status record. The Phase 5 retention / Phase 6 deletion decision
for Design A is preserved, not flattened.
clients/ un-ignored and tracked, including the .service unit and README:
deployed code belongs in version control. Design A is NOT deleted here.
progress.md (finding 2): the file was deleted after 636ed8a, yet CLAUDE.md
instructed every session to cross-check against it. References removed from
CLAUDE.md, AGENTS.md, internal/orchestrator/rotation_test.go (comment only)
and deploy/hooks/orchestra-codex-poll.sh; AUDIT.md now carries the log role.
web/go.mod (finding 4): a module stub ends the parent package graph at the
directory boundary, so go list ./... no longer yields
web/node_modules/flatted/golang/pkg/flatted. A build tag cannot work - the
package is in the package list before tags are evaluated. Local/CI-only
breakage: Dockerfile.api builds ./cmd/orchestra by explicit path and
.dockerignore already excluded node_modules.
orchestra-worker (finding 5): untracked (8.9MB, mode 100755, still on disk);
both binaries now gitignored.
Token compare (finding 7): cmd/orchestra/main.go:139,582 use
subtle.ConstantTimeCompare, matching the authz.go idiom. The token != ""
guard stays first, so an empty configured token still means auth-disabled
rather than auth-bypass. Three further plain != secret compares remain in
internal/federation/federation.go:343,346,368 - tracked, not fixed here.
Also included from the review pass: orchestrator.go records adapter-resolution
failures in SessionHealth.LastError instead of dropping them on a bare
continue, plus an Observed flag so lease-seeded health is not mistaken for a
live reading, with a covering test. GET /v1/tasks/<id>/health now returns a
record with last_error where it previously returned a bare 404.
REVIEW.md's own second pass claimed every checkable fact held up; four did
not. AUDIT.md never contained the false Design B claim (AGENTS.md was the
second copy), the guardrail is at orchestrator.go:312 not :309, the
progress.md site list missed the codex-poll hook, and only orchestra-worker
was tracked. Verified: go build, go vet, go test, and
go list ./... | grep node_modules all clean with every change applied
together. No live herdr or pane was touched; nothing was deployed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GEugbHVYfAXFpTqDYbByEB
All five defects filed while implementing B18, plus the router flake that
predated them. None of this has run on the deployed instance: the service
is stopped and /usr/local/bin/orchestra predates every change here.
B20 is the one that could silently defeat approvals. The capture revision
was UnixNano, so it changed on every read and said nothing about whether
the pane had changed; it is now an FNV-1a hash of the pane text, changing
iff the text does. The worse half was precedence: capture() preferred the
coordinator over a published worker capture, handing Queue a timestamp the
owning worker's staleness check could never match, so every federated
approval resolved "stale" and the keystroke never happened. Worker captures
now win — their existence means a registered worker owns that pane — and
capturePane follows the same precedence via Capture.Source rather than
guessing.
B19 was filed as "federated approvals emit no event", which overstated it:
the resolution half already existed, and correctly fires only on an
acknowledged worker report. The missing half was the request. Server.action
now appends ApprovalRequested at queue time, subject_ref set to the command
ID the later resolution carries. If that append fails the queued command is
resolved "rejected" — a keystroke that left no audit trail must not run.
B21 bounds the command list: resolved commands prune after 30 minutes on
both Queue and Commands, pending ones never at any age, since dropping one
would discard an operator decision. The persistence half stays open and is
recorded as such — captures and commands are still in-memory only.
S12 splits ORCHESTRA_NTFY_TOKEN, which was both the secret handed to the
ntfy server and a valid inbound credential for the ntfy surface; the latter
is now ORCHESTRA_NTFY_SURFACE_TOKEN. Breaking: a deployment relying on the
old dual use has no inbound gate until it sets the new variable. S13
deletes the dead auth() copy of the authorization policy.
The router flake was in the test, not in assignment. Store.Tasks() ranges a
map, and the assertion indexed two separate Tasks() calls, failing whenever
the orderings disagreed; instrumenting it showed a valid TaskLeased and a
genuinely leased task on every "failing" run. It now snapshots once and
asserts that exactly one task is leased, and passes at -count=60.
AUDIT.md records what is still not done: the deployed env and binary, the
live re-verification B13-B17 has always lacked, and two operational faults
found in the journal that block it — all six herdrs are refusing
connections, and ntfy delivery is failing 403 on every send.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01535A3Y8RtkAi8wYuWhtkEd
The UI is a full control plane: it can create tasks, release or complete
them, and inject approval keystrokes into live panes. authz.HTTP did cover
it (an absent surface header defaults to Web), but the gate is opt-in and
the deployed env sets no tokens while binding all interfaces, so in
practice it was reachable unauthenticated from the LAN. Setting the token
alone did not work either: a browser cannot put a bearer token on a
document load, so the UI would 401 on index.html.
- ORCHESTRA_WEB_TOKEN is now mandatory; startup fails rather than silently
serving an open control plane.
- authz.Sessions issues random values stored SHA-256-hashed, with a TTL,
so a leaked snapshot yields nothing usable.
- POST /v1/ui/session verifies the token in constant time and returns it
as an HttpOnly, SameSite=Strict, Secure cookie. This is a presentable
form of the same credential, not a new authority.
- HTTPWithSessions accepts that cookie in place of the bearer token, and
only for the Web surface. The login endpoint and non-/v1/ GETs (the SPA
shell) are exempt by necessity; every /v1/ control path stays gated.
Note this is a breaking config change: .orchestra-config/orchestra.env
sets no tokens, so the service will not start until it does, and setting a
Web token newly gates the other /v1/ surfaces that default to Web.
AUDIT.md is reconciled against the code rather than against itself. B14,
B15 and B16 are closed with their evidence; B17 is closed on the worker
path only; the stale claim that B13 was open is corrected. Adds the
previously undocumented command channel and web UI, and files what that
implementation pass surfaced: federated approvals emit no event (B19), the
local capture revision is a timestamp rather than a change counter and can
silently defeat approvals (B20), the command queue never prunes (B21), the
ntfy token serves two unrelated purposes (S12), and a dead copy of the
authorization policy sits in main.go (S13).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01535A3Y8RtkAi8wYuWhtkEd
Introduces the browser-facing surface and the worker-side protocol that
backs it:
- internal/ui: joined read model plus per-task lifecycle and approval
controls, kept separate from the raw endpoints workers and harnesses
depend on.
- internal/webui + web/: Vite/React app, build output embedded via
go:embed and served as an SPA fallback.
- federation: per-(worker, task) captures with a monotonic revision that
advances only when pane text actually changes, and a command queue
restricted to grant_approval / deny_approval, each bound to the capture
revision the operator acted on.
- orchestra-worker: publishes captures and executes commands only after
re-reading the pane and confirming the revision still matches. Sends
keystrokes only for a visible y/n prompt or OpenCode's fully labelled
selector, and refuses to deny through that selector rather than guess
at unobservable navigation.
This is the ownership boundary AUDIT.md's B14 and B17 call for: approval
becomes an explicit, revision-bound operation executed by the worker that
owns the pane, instead of a side effect of prompting over a
coordinator-driven remote socket.
Also ignores the web build inputs and outputs. node_modules ships vendored
Go packages, so go build and go test walk into it if it is merely
untracked; both node_modules and .node_modules are excluded.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01535A3Y8RtkAi8wYuWhtkEd
kami
kami