A per-task page at #/tasks/{id} with tabs for the attempt's stages and log, changed
files, configuration and a raw debug view. Reachable from the task list, deep
linkable via ?tab= and ?attempt=, and reload safe. Existing URLs (#/tasks?task=N
and #/tasks?status=...) are unchanged.
- A stage rail per attempt: name, duration, terminal status and error text, with
unattributed events surfaced and the agent's own verdict shown separately.
- No pass/fail badge and no exit code anywhere: the backend has neither, and a stage
that returned nil after producing a poor diff is a success. Status word, duration
and error text are shown instead.
- The log pane filters by attempt, stage and level, states the coverage caveat for
stage filters (agent and tool output records no stage), and keeps the three read
outcomes distinct: no log reader in this deployment, no file for this attempt, and
no entries matching the filter.
- Retry is labelled as a new run and its confirm says the previous attempt's commits
are discarded, because it hard-resets the branch onto the base.
- An absent capture reads "not captured for this run" rather than implying nothing
changed, and no elapsed-run figure is shown since only per-stage durations exist.
- The events cache is invalidated on retry (it previously kept the ended attempt's
events) and the log cache is keyed by attempt, so two attempts cannot be conflated.
- ui/dist is rebuilt and committed: the gate runs the node tests but never rebuilds
the bundle, so a stale one would otherwise ship behind a green gate.
- docs/prds/task-run-detail.md records the surface and its limits, and
docs/architecture/observability.md describes it alongside the read endpoints.
Both new vocabularies are pinned by tests on each side (config schema, change
statuses) so a one-sided rename fails a test instead of silently degrading the page.
The events table had no attempt column, so two attempts of the same task were
indistinguishable: a retried task's stage events interleaved into one timeline and
nothing could say which run produced a given result.
- events gains an `attempt` column through the existing presence-gated
migrateTasks loop rather than a version bump, so user_version stays 3 and an
existing archie.db is migrated in place. A legacy-database test builds the
pre-change schema and proves both that existing rows survive unattributed
(attempt = 0) and that a new value round-trips.
- Event.attempt = 11 and ReadTaskLogRequest.stage = 9 are field additions to
existing messages: no new RPC, no new table, and no change to any canonical error
message that crosses the state-store wire.
- Every task event the workflow engine emits now carries the attempt. Producers
that have no attempt yet (task_queued, binding_dispatch_failure) stay
unattributed rather than guessing one.
- Changed files are captured as a durable changes_captured event at commit and push
time, reusing the existing three-dot diff primitive: paths from FilePatches(),
counts from Stats(), base_sha recorded as the merge base actually diffed rather
than the base tip, a 200-file cap with totals kept complete. Consumption goes
through an unexported optional interface so workflow.Trees -- the yaegi capability
surface for interpreted stage plugins -- is not widened. A capture is also taken
once the pull request exists, so the capture can link it.
- The per-attempt configuration is captured as a durable config_captured event
carrying the non-secret config.TaskConfig, instead of a new table plus two new
state-store RPCs: events are already this system's per-attempt provenance stream,
so a table would have given one feature two provenance mechanisms. Redaction is
proven against canary secrets. Measured payload: 539 bytes typical, 2333 bytes at
20 models.
- Captures carry a schema marker and are refused unless their key set matches the
declared payload shape, so drift surfaces as "recorded but could not be read"
rather than as a capture of zeroes.
- Three read endpoints serve the attempt rail, the changed-file captures and a raw
debug view, mapping to UI-owned response types rather than dumping store types.
- The task-log read gains a stage filter, bound to the stage logger for that
stage's own lines; agent and tool output records no stage.
A task-scoped grant may already insert events on its own task, so this provenance
is exactly as trustworthy as the rest of that task's event stream. That limit is
documented rather than papered over.
Add two field-only extensions to the State Store contract, no new message
and no new RPC: Event.attempt (11) gives every task event an attempt key,
and ReadTaskLogRequest.stage (9) lets a log read be scoped to one stage.
Generated stubs regenerated with tools/proto.sh and committed as LAW so
proto:check, which diffs the working tree against the index, can pass.
Two fixes to the published site, both about not shipping something false.
exclude_docs held back prds/ but let ui-service-boundary.md through, because
architecture/migration-decisions.md links to it three times. That PRD in turn
links to service-decomposition.md and config-schema.md, so the exclusion was
still producing two dead links on the published site. Letting those two
through closes the chain: the build now reports no excluded-link references
at all.
generated-documentation.md's "Step 9: CI" said no workflow builds, checks or
publishes documentation. Adding .github/workflows/docs.yml made the first
clause false. Restated to say what that workflow does and, more importantly,
what it deliberately does not: it does not run docsgen, regenerate
contracts.json, or publish. Generated-artifact drift is still gated locally by
task check, which is the split that section exists to protect, since a CI job
that regenerated output could paper over uncommitted differences rather than
fail on them.
Build-only. The site deploys from the Ansible control node via the
vectislabs.automation docs_site role against nebula, so CI has nothing to
publish; what it catches is a docs change that will not build at all, before
it reaches a deploy.
Path-filtered to docs sources, so the ~600 commits a month that touch no
documentation do not run it.
Adds an mkdocs.yml, a landing page, and `task docs:site` / `task docs:serve`.
Deliberately hosting-agnostic: MkDocs is the same input Backstage TechDocs
consumes, so choosing where this lands later costs nothing.
exclude_docs holds back four subtrees, since everything under docs/ would
otherwise be published. archive/ is superseded design and would compete with
the current architecture for a reader's attention; inspiration/ discusses
other products by name; prds/ are open decisions rather than settled design;
github-token.md is credential setup. One PRD is let back through because
architecture/migration-decisions.md links to it three times, and a page
excluded out from under a live link is just a 404 with extra steps.
No nav block. MkDocs derives it from the directory layout, so a new page
under docs/architecture/ publishes without a second edit, and there is no
hand-maintained tree to drift against the repository.
Builds --strict, so a dead internal link fails rather than ships.
Add `task deadcode` (whole-program RTA from ./cmd/...) and `task reachaudit`
(the existing classifier on top of it). Neither joins `task check`: both
report a standing backlog, so gating on them fails every run until that
backlog is zero, same as `task vuln`.
`task deadcode` shells to `go run ...@latest` rather than a local binary on
purpose. A deadcode built against an older Go than this module exits non-zero
and prints nothing, which reads as "no dead code" instead of "analysis
failed" -- the locally installed copy was built against 1.26 against a 1.27
module and had been silently reporting an empty tree.
reachaudit had two traps that made it answer confidently from nothing:
its --repo default pointed at a path that does not exist on either machine,
and it reused /tmp/reachaudit-deadcode.txt with no staleness check, so it
would serve a previous tree's analysis indefinitely. Key the cache on HEAD
and refuse it outright when the tree is dirty.
Also lands the design for tools/symbolrefs, the type-resolved reference
index that covers what reachability cannot: RTA treats every address-taken
function as live, so every reflect.ValueOf entry in the *extract tables
launders a dead symbol into a reachable-looking one, and deadcode reports
functions only, leaving dead package vars unrepresentable.
All three Greptile findings were valid against the code; each is fixed at its
cause rather than at the symptom, and each fix is covered by a test that
fails when the fix is reverted.
**Check could silently write.** `check` was matched only as the first
argument, and leftover positional arguments were never inspected. So
`docsgen --repo-root .. check` — and any misspelled subcommand — fell
through to the write path and overwrote the artifact the check promises not
to touch. That is the exact guarantee the mode exists to provide, so the
failure was silent and in the wrong direction.
Extracted argument handling into a testable `parseOptions` and made a leftover
positional argument an error with usage text. Switched the flag set from
`ExitOnError` to `ContinueOnError` so the error surfaces as a value rather
than a process exit, which is what makes it testable.
**Obsolete detection claimed files it does not own.** It treated every
sibling `.json` as docsgen's and reported it obsolete, so checking an
artifact in a mixed-use directory produced a false stale result, and the
documented target layout — which anticipates further named artifacts in
`docs/data/generated/` (catalog.json, AsyncAPI output) — would have been
reported stale as soon as a second artifact legitimately existed.
Ownership is now declared rather than inferred: `generatedArtifacts` maps each
directory docsgen owns end-to-end to the names it produces there, and a
directory docsgen does not own is never judged. A new unit test pins both
halves — the orphan inside an owned directory is still caught, and a foreign
`.json` in an arbitrary `--out` directory is not.
**Provenance traversal stopped at one level.** Refs were walked only inside
the top-level contract schemas, so a definition reachable solely through
another definition resolved to "no authoritative definition found" — the
diagnostic lying about the code it points at. Replaced the single pass with a
breadth-first walk that follows discovered definitions transitively, seeded
with every published contract first so a definition reachable from several
contracts is attributed to the alphabetically first rather than to whichever
the walk happened to reach. `walkSchemaRefs` now iterates map keys in sorted
order for the same reason.
`schemaProvenance` takes its roots as a parameter, which is what makes the
transitive case testable without real types.
Verified by mutation: with each fix reverted in turn, its test fails —
`parseOptions(["--repo-root" ".." "check"]) returned nil error`,
`obsoleteArtifacts() = [...someone-elses.json is obsolete...]`, and
`provenance[Leaf] = ""`. Control run passes unmutated, and the file is
byte-identical after restoration.
Also confirmed end to end that the previously silent invocation now exits
non-zero with usage instead of writing, and that the committed artifact is
untouched by it.
`task check`: exit 0, 0 failures.
Closes#855 (archie-core-5gzx).
The check mode added in the preceding commit had no caller, so it prevented nothing: a
stale contracts.json still passed every gate. This wires it in and makes
drift fail the build.
- `task docs:generate` writes the committed artifact.
- `task docs:check` verifies it, non-destructively.
- `task check` runs `docs:check`, right after `proto:check`, so both
generated-artifact surfaces are judged the same way in the same place.
The check cannot be satisfied by rewriting: it generates into a temporary
directory, so a failing gate leaves the artifact exactly as it found it and
the author applies the regenerated bytes. That keeps the regenerated diff
reviewable in the same commit as the code change that caused it, which is
what Step 9 of the architecture document requires and what a CI auto-update
would quietly bypass.
Documents that called this gap open are updated rather than left to
contradict the code:
- `generated-documentation.md` Step 8 (no `docs:*` task exists -> both exist),
Step 9 (drift is gated locally, not in CI), and the hook/CI policy (rule 3
is now enforced by the gate; rules 1, 2, and 4 still rely on the author,
since nothing observes a commit).
- `contract-conformance-audit.md` gap 1, now closed.
- `config-example-drift.md` item 3, which asked for exactly this wiring.
- `CLAUDE.md`, whose `task check` description listed neither `docs:check`
nor `test:tools`.
- The build, docs, and diagnostics skills, plus a re-baselined diagnostics
snapshot: the `task_docs` delivery surface went 0 -> 6 and the `tools`
module 4 -> 7 now that the tasks exist.
Verified with `task check` (green), including the new `docs:check` step.
`docs/data/generated/contracts.json` sat stale since a825db28 with nothing
to notice: no task, no test, and no CI step regenerates or compares it, so
the committed contract reference quietly drifted from the Go types it
claims to describe. It was only found by hand-running the generator.
Implements Step 6 of docs/architecture/generated-documentation.md, which
already specified this command and closed the gap filed in
docs/prds/contract-conformance-audit.md ("`docsgen check` is documented and
does not exist"):
1. generates into a temporary directory;
2. normalizes line endings;
3. compares the schema trees with the committed output;
4. reports missing, changed, and obsolete files;
5. names the authoritative definition behind each mismatch;
6. exits non-zero without modifying the working tree.
The check never writes to the working tree, so the author applies the
regenerated bytes rather than the gate silently doing it. That is why this
is a `check` and not a CI auto-update: the code change that shifted the
contract surface should be reviewed with its regenerated artifact in one
commit, and a bot commit would bless an unintended contract change with no
reviewer ever seeing it. Step 9 states the same prohibition directly.
Diagnostics name the Go type, so a failure says where to look:
docs/data/generated/contracts.json is stale; regenerate it with:
go -C tools run ./docsgen --repo-root ..
schema "AgentExecutionResult" changed (authoritative definition:
github.com/samcharles93/archie-core/internal/agentexec.Result)
schema "MediaAttachment" changed (authoritative definition:
reached from .../internal/gateway.MessageEvent)
Nested definitions are attributed through the published contract that
references them, iterated in sorted order so the diagnostic is
deterministic when several contracts share a definition.
Usage: `go -C tools run ./docsgen check --repo-root ..`, plus the existing
`--out` to compare an arbitrary artifact.
Reuses the write path rather than reimplementing encoding, so check and
generate cannot disagree about the expected bytes.
Tests cover: the committed artifact is current (the regression guard that
would have caught this), changed/removed/missing/obsolete detection, and
that a failing check leaves the artifact's bytes, mode, and directory
contents untouched.
`task docs:check` wiring into `task check` is deliberately not in this
commit: Taskfile.yml is being edited concurrently in this working tree.
Verified with `task check` (green).
All ten `archie-*` skills carried claims that no longer matched the
checkout. Re-verified each against code, git history, and the shipped
configuration, then corrected the drift.
Structural changes the skills had not caught up with:
- Composition root moved: `cmd/archied/main.go` is a ~30-line dispatcher
into `internal/app/archied.Run()` with wiring in `bootstrap.go`. The
skills still described `cmd/archied` as the wiring site, which also
silently zeroed the diagnostics delivery snapshot.
- Config decoding moved out of `internal/config` into
`internal/infrastructure/configuration`. `config.Load` / `config.LoadDir`
no longer exist; the loader is `configuration.New` + `Loader.Resolve`.
- Task contract moved to `internal/domain/storecontract`, with
`internal/store/interface.go` holding aliases only. The task type is
`internal/domain/workflow/task.Task`.
- `internal/workflow` became `internal/domain/workflow`; `internal/storerpc`
became gRPC `internal/infrastructure/staterpc`; `internal/policy` does not
exist.
- `Store.Transition` now guards on `from` and writes status plus history in
one transaction. Three skills still described the old non-atomic behavior
as current and open.
- `internal/domain/{messaging,workflow,workintake}`, `internal/eventbus`,
`internal/app`, and `internal/infrastructure` now exist.
`internal/domain/{agent,identity}` and `internal/policy` remain target-only.
Tooling and environment facts:
- Go 1.27.0 (root and tools modules), golangci-lint 2.13.2, staticcheck
2026.2.1; snapshot dates and HEAD moved to 2026-09-18 / 0ee2d27e.
- Re-ran the bundled diagnostics and re-baselined its snapshot table.
Confirmed 78.6% store coverage, a clean race run, and zero dupl findings.
- Corrected the claim that no `.dockerignore` exists; it does.
- Replaced skill script paths that pointed at the untracked `.claude/skills`
symlink with the canonical tracked `.agents/skills` location.
- Fixed the delivery-snapshot comment's wrong commit hash for the Gitea
workflow removal (dd9bddc2, not e539048).
Findings the verification closed:
- The VitePress site and `.github/workflows/docs.yml` were removed
(de150f20); the Gitea workflow was dropped (dd9bddc2).
- The changelog `vX.Y.Z`-vs-`X.Y.Z` contradiction is settled, not open: both
changelogs and the CI gate agree on `## [<version>]`.
- Tracked `docs/node_modules` and the skill-script test regression are
settled, not current.
- The six unmerged `origin/fix/*` and `origin/feat/*` candidate branches no
longer exist.
Also regenerates `docs/data/generated/contracts.json`, which the
verification found stale since a825db28: docsgen emits 11 schemas now, and
`MediaAttachment.raw` plus an `AgentExecutionResult.required` field were
missing. `cmp` passes after regeneration.
Verified with `task check` (green).
The agent-side half of the same gap. A stage failure already logged
"stage failed -- parking" before parking, but a run whose OUTCOME was parked
(a review that could not run, a workflow that ended with no outcome) reached
the timeline as a KindParked event and the store as park_reason and never
the log an operator downloads.
park is the one place a park's reason is final, so it is now the one place
that logs it; the stage-failure branch no longer logs its own near-copy, and
the stage names itself inside the reason. The entry carries the stage as a
field so nothing is lost by dropping the duplicate line.
observability.md's "Implemented design" section documented the container's
system-log path as though it were the only source of task-log content; it
now records the daemon's and the workflow's own park writers, including why
the entry follows the transition rather than preceding it.
An operator who asks "why did task 1 attempt 3 park?" and downloads that
attempt's log got a zero-byte file, because nothing daemon-side wrote to a
task log at all: the container's system-log subscription was the only
writer, so an attempt that parked before any container produced output --
a worktree prepare failure, an unavailable worker, a taskrun error -- left
the file the sink had created empty and the reason only in the daemon-wide
log.
Both park paths now write the reason into that attempt's sink. The entry is
written after the guarded transition lands, so a park another actor won
does not leave a log claiming an outcome the task board contradicts -- the
same race the parked event is already withheld for.
Write built its own context.Background() for the handler call, which meant
any new caller with a context to offer could not hand it over. Taking the
caller's context is what the next commit's park-reason writer needs, and it
is what contextcheck requires of a daemon-side writer.
No behaviour change: the context is used for the handler's own attribute
extraction and never consulted for cancellation -- a park is recorded at
exactly the moment a run's context has been cancelled, so a cancelled
context must not suppress the entry.
A container run only restored ownership of the bind-mounted worktree on the
CommitAll/Push success path (archie-core#520). The agent runs as root inside
the container, so every other way a run can end -- a gate failure that parks
the task, a stage error, an early return -- left the worktree owned by UID 0
on the host. The daemon is the process that later cleans and resets that
directory as its own non-root user, and it cannot even unlink inside a
root-owned directory, so the next attempt failed before it could start no
matter what the worktree refresh did.
runTask now defers the same reconcileOwnership the commit and push paths
already call, so it runs at every exit. The retry recoveries in worktree
package cannot fix this from the daemon side: a non-root process cannot
chown, and removing the directory needs write access to it.
Observed through a supplementary group -- a non-root process may move a file
it owns into a group it is a member of -- so the test proves the chown ran
without needing root.
Residual, filed as archie-core-zcg8: a container that is SIGKILLed never
runs this defer.
A retry after any attempt that did not commit failed with
worktree prepare failed: checkout branch <branch>:
a branch named "refs/heads/<branch>" already exists
on a branch that was never the problem. refresh inferred "the branch does
not exist yet" from a failed checkout and reported whatever the Create
fallback then returned, so the real failure was replaced by a false one --
and the operator had no way to tell that the park reason was fiction.
The real failure, reproduced: a stage running in the container can leave a
path whose TYPE contradicts the tree refresh resets onto -- a directory
where the base tree has a file, or a file where the base tree has a
directory. go-git's hard reset cannot write over that contradiction and
fails with `openat <path>: is a directory` / `not a directory`, which is
exactly the failure the Create fallback then relabelled. Retry could not
recover from it, and a second path of the same class (a branch removed
while HEAD still pointed at it) failed with "reference not found".
refresh now clears the worktree against the tree it is resetting ONTO
before it checks anything out, which removes exactly those conflicts -- an
on-disk directory the tree declares as a file is untracked, and so is a
file the tree declares as a directory -- and resolves the branch's
existence from the reference store instead of inferring it. Resume gets the
same treatment, so the remediate path cannot fail the same way.
This is the third fix to this function (Force in 05944b6f was the second),
so the checkout/reset/clean sequence is rewritten rather than patched again.
Every client method that decodes a repeated proto field goes through
mapValues, which returned nil for nil input. Protobuf decodes an empty
repeated field to nil, so an empty collection crossed the wire as JSON null
rather than an empty list: the field's type depended on whether it happened
to have contents, and the wire path disagreed with the local path, which
normalises its own slices.
This is the nil-slice class the contract-conformance audit deferred, arriving
by the route its own revisit condition names -- a handler that can produce
nil. The audit checked the webui handlers (clean, as it said) and did not
cover the wire mapping added afterwards, so the regressed field was the new
one: ReadTaskLog's components, observed as "components": null from the live
dashboard.
mapValues now returns []B{} for empty input, which fixes the class once at
the mapping helper rather than at each of its 36 call sites -- none of them
compares the result to nil, so nothing depended on the old behaviour.
TaskLog also normalises components, which is read straight from the proto
rather than through mapValues.
Red first: the test drives the existing local/grpc matrix and asserts the
marshalled JSON, so it fails on a null collection without depending on which
path is under test. Both collections are exercised empty -- entries present
with components empty, and a log file that exists with no lines -- because
the first shape alone leaves the entries side untested. Verified by mutation:
restoring the nil return fails it on "entries marshalled to null".
The download sets its attachment headers before the read, because the body
streams straight from the reader and a log is unbounded input -- buffering it
to size the response would defeat the reason it streams. Nothing then cleared
those headers, so a 404 for an attempt with no log went out as
Content-Disposition: attachment; filename="task-1-attempt-0.log"
Content-Type: application/x-ndjson
A browser treats that as a download: the caller gets a .log file containing
the text "no log recorded for this attempt". A stale or direct download URL is
enough to trigger it. The 503 for a process with no reader had the same shape.
Both error paths now clear the attachment headers before writing, so the
response is an error rather than a file. The successful path is unchanged.
Red first: the existing test asserted the status only, which is how this
passed. It now asserts the headers are cleared on a 404 and on a 503.
The attempt-log panel offered no way to get a log out of the dashboard, and
its failure copy asserted that task logging was not enabled for the run. The
first is the feature: the panel now links to the download route for the attempt
it is showing, so the browser owns the filename the server sets and streams the
body to disk rather than the client holding a whole log in memory.
The second is the honesty fix. "No log recorded for this attempt" is a fact
about the attempt; "this dashboard cannot read task logs" is a fact about the
deployment, and only the second is worth sending an operator to a setting for.
The panel renders them differently off the response's found/disabled markers.
Rebuilds ui/dist, which is a committed asset.
The dashboard's attempt-log panel answered every read with "task logging is
optional and was not enabled for this run". That message was false: logging is
wired unconditionally in the daemon, and the log files exist. The dashboard
process simply had no reader -- composition left TaskLogs absent because no
contract carried it -- so the handler saw a nil registry and reported the
disabled shape, which the page rendered as a claim about configuration.
internal/logging now answers whether an attempt has a log at all: Tail sets
Result.Found, so "this attempt has no log" is a value rather than an empty
list indistinguishable from a read that never happened. Its registry gains
TaskLog (a decoded page) and TaskLogContent (the file verbatim, streamed,
because a log is unbounded input), and an absent reader reports
ErrTaskLogsUnavailable rather than an empty page -- a process that cannot read
logs must not implicate the attempt.
Those two answers travel the State Store contract as storecontract.TaskLogStore
(ReadTaskLog, StreamTaskLogContent) and are rehydrated on the client by
(Unavailable, canonical message), so archie-ui composes the reader from its own
store client instead of opening another process's files. The daemon's registry
is wired as the server-side reader in the State Store process only.
On top of that read, GET /api/tasks/{id}/logs/download serves the attempt's log
as a Content-Disposition attachment named task-<id>-attempt-<n>.log, and the
panel offers it. The JSON route keeps its shape and gains found, so the page
states the attempt has no log instead of blaming a setting.
Task log files stay owned by their directory's owner: archiving still removes a
task's log directory, and that removal is the daemon's, not the dashboard's.
One task attempt's persisted log lives in the state directory the State Store
process owns, and no contract carried it: the dashboard process therefore had
no way to read a log at all, and answered every attempt with "task logging is
optional and was not enabled for this run" in a deployment where logging is
unconditional.
ReadTaskLog returns a decoded page and StreamTaskLogContent returns the raw
bytes server-streamed -- a log rotates at 20 MB, past gRPC's unary cap, the
same reason StreamCaptures superseded ListCaptures. found=false with no error
means the attempt has no log; a service with no reader answers Unavailable,
and keeping those apart is the point: only the second is a deployment
condition.
Slice 1 of docs/prds/memory-engine-unification.md: persist the
participant, so a later curator can address agent-user memory scope.
- messages gains sender_id: schema, ALTER migration (idempotent via
duplicate-column detection), both inserts, and both history selects.
Pre-existing rows default to empty.
- curator.SessionSummary gains AgentID and ConversationMessage gains
SenderID; the sessioncurator adapter fills them from the session's bot
identity and the stored record.
The memory contract is untouched and nothing consumes the new fields
yet. Deriving a session's participant from its user-role messages, and
failing closed on ambiguity, stays with the curator rewrite.
Targets are deduplicated by path so a shared filesystem is probed once, but
the first target seen won. With root, then an optional /var, then a required
data path that resolves to /var, the optional target is the one that survived
-- so its rules were applied to a path the daemon requires. An absent /var is
skipped as "optional absent", and with another healthy target present the
probe reported OK while a required path was unavailable.
Deduplication now keeps the strictest requirement any target places on a
path, and its name, so the failure is both visible and correctly attributed
rather than reported against the optional target that happened to come first.
Red first: the new test drives the reported scenario through an injected
statfs and failed with `status = "ok", want degraded` before the fix.
Verified by mutation -- restoring first-wins dedup fails it again.
Two defects in the activity/answer budget, both found in review.
The tool block returned as soon as the newest line alone fit, so a turn with
several short tool calls discarded history that was never over budget and
reported a "+N earlier" that was untrue -- those lines had room. The window
now widens from the newest line backwards while the result still fits, so
every line that fits is kept. The indicator appears only when something is
genuinely left out.
The opening frame sliced the answer to its budget without an ellipsis, so a
large first chunk -- or a turn that never opened a reply, leaving messageID 0
-- rendered text that appeared to begin mid-sentence, while later frames
marked the same cut. Telegram messages are immutable, so it could never be
corrected. Every frame now marks the cut; the special case only ever bought
one rune.
TestLiveReplyFramesStayWithinOneMessage asserted the ellipsis on the last
frame only, which is what let the unmarked opening frame through. It now
checks every frame, and the count bound follows: identical frames are
deduplicated by render(), so an unmarked opening frame was previously a
distinct body where a marked one is not. Verified by mutation -- restoring an
unmarked opening cut fails the test on frame 0.
Rebased onto main, which now carries an equivalent STATE_STORE_TOKEN isolation
guard in state_store_client_test.go from an independent baseline repair. The
two changes were the same fix expressed twice, so main's wording is kept and
this branch's duplicate is dropped -- the file is now identical to main and no
longer appears in the diff.
That leaves one commit, touching only internal/channels/telegram/live.go.
Removing clampToOneMessage, which this change orphaned: all three of its call
sites were replaced with clampToRunes directly, and `unused` flagged it. Dead
code left behind by a rewrite does not ship.
TestShutdownWatchdogHangingPastLeashForcesExit gave the watchdog two leash
lengths (60ms) to force an exit, but the 30ms leash only starts the timer --
the firing goroutine must then be scheduled before the injected exit is
observable. Measured margin is ~29.7ms, so the test is a wall-clock race
against the Go scheduler and fails intermittently on a loaded machine, where
it parks real work behind a gate failure unrelated to the change under test.
The assertion is now the contract that actually matters and that scheduling
jitter cannot falsify: the exit must not happen before the leash, and it must
happen. Waiting for it uses a generous ceiling, and the new elapsed >= leash
check catches an early fire strictly -- something the old test could not do,
because it only bounded lateness.
The inverse test asserts the ABSENCE of a forced exit, which load cannot
falsify (a slow runner only delays the exit being checked for), so it keeps a
tight bound and the suite stays fast.
Verified by mutation: removing the forced exit fails the first test, and
firing before the leash now fails it on the new elapsed check.
Review found that a bare "showing 3/197 lines" inside ordinary output was
accepted as truncation metadata and replaced by a summary, hiding the real
result. Both built-in producers wrap the notice as "[truncated: …]"
(internal/tools/builtin/truncate.go), so the wrapper is now required rather
than tolerated: it is the only marker separating a notice the tool printed
from content that merely mentions a count.
Four findings from review, all valid:
1. Search detection is now gated on the tool name rather than applied to every
result. A compiler diagnostic ("file.go:42: message") has the same shape as
a search hit, so counting it reported a build error as "N matches" and hid
the diagnostic.
2. Only ripgrep match lines are counted. The hyphen context form
("path-N-text") and "--" group separators were being counted as matches, so
a search with context requested reported the match count plus every
surrounding context line.
3. The notice parser accepts shell's "showing last N/M lines" form as well as
the plain "showing N/M lines" one. Shell emits the "last" form, which failed
integer parsing and was not recognised as metadata -- so a truncated shell
result quoted the raw truncation notice as its preview, the exact defect this
change set out to remove. The summary now also states which end was kept
("97 lines total; showing last 66" for a tail truncation).
4. The renderer cases are consolidated into one table-driven test per the
development protocol in AGENTS.md, covering all four defects above plus
singular/plural, colons in paths, and module-cache noise.
Operator report: tool output in the Telegram channel was unreadable during
search-heavy turns -- a wall of fenced blocks each showing one arbitrary
match line ("/path/config.json-322- }") under a heading that said
only "grep -- done".
The renderer quoted the first non-module-cache line of a result and appended
a line count recomputed from its own truncated sample, which could not agree
with the total the tool had already reported: one search rendered the tool's
"194 more lines" beside our "… 3 more lines". Two different numbers for one
result set read as noise, and the quoted line carried no context at all.
- Describe, don't quote: a search is reported as "N matches", a truncated
result as "<total> lines total; showing <n>", and every other tool falls
back to its first informative line. Truncation notices are skipped, since
they are metadata rather than content.
- Never invent a count: the tool's own reported total is the only
trustworthy source, parsed from its truncation notice rather than
recomputed from the visible sample.
- Remove the now-dead compactToolPreview and countNonEmptyOrContentLines.
Also bump go-telegram/bot v1.25.0 -> v1.27.0, which fixes a streaming bug
that lands squarely on this path: rawRequest sent the multipart body through
an io.Pipe, which net/http cannot replay, so sendMessage/editMessageText
failed until the connection was dropped.
Remaining work from the report (collapsing repeated calls into one counted
entry, and a ring of recent activity lines with a capped frame) is tracked
in #837.
The zip and its Release were undocumented: RELEASING.md described the
images and the gate, and nothing about dist-zip, the Release asset, or
that a gateway-only release does not rebuild the agent image. Adds that,
plus the consequence that bit this release -- Actions runs the workflow
from the tagged commit, so a deploy.yml fix must be an ancestor of the
tag rather than a commit on top of it.
The dist-zip job built the linux/amd64 zip and uploaded it as an Actions
artifact, then stopped. An artifact is reachable only from the run's page,
behind an Actions login, and expires after 90 days -- so every release since
the zip job landed produced a download an operator could not find. The one
Release on this repo (archied/v1.22.0) was created by hand and carries no
asset at all: nothing in CI has ever created one.
Publish the Release from this job and attach the zip:
- Job-level `permissions: contents: write`. The workflow default is
contents:read, so a publish step would fail without it; scoping it to this
job keeps the build-and-push jobs read-only.
- The body is this version's changelog section, extracted from the "## [x.y.z]"
heading to the next "## [" heading rather than pasting the whole file. The
changelog stays the single source of truth, and the tag/changelog gate in
`gate` has already proven the section exists.
- Attached to the archied tag, because the zip is named and versioned for
archied; `--follow-tags` pushes both tags at one commit, so the runtime
release is the same commit.
- Idempotent: `gh release view` decides between `create` and
`upload --clobber`, so a retried or superseded deploy run cannot fail on an
existing Release or duplicate the asset.
- Fails loudly when the section extracts empty, rather than publishing a
Release with no notes.
task fmt reformatted this file's fmt.Fprintf call; committing it rather than
letting it ride in the release commit, so the release diff stays the changelogs
alone. AGENTS.md: formatter changes are adopted verbatim.
Approved for implementation, dated 2026-09-16, previously untracked.
Settles four open items in docs/architecture/migration-decisions.md#5 and
supersedes the memory-turn-context.md draft.
Section 8's deletion list no longer names config.Memory.SessionTTL: the
council carried that deletion standalone and af3a8a2 removed the field,
so listing it here would schedule work already done. Provider and
ProviderConfig are unchanged, both still present.