A session can exit and run onExit (which deletes c.terms[id]) in the gap
between subscribe returning exited=false and openTerminal assigning
c.terms[id]. The delete is a no-op there since the key isn't set yet, so
the later assign resurrects a stale entry for a dead pane, trapping every
future open for that id on the connection. Re-apply the delete after the
assign when onExit fired in the window, tracked by a c.mu-guarded flag.
Add a stress regression test that races the exit against the assign.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The exit callback enqueued the exited frame before deleting c.terms[id],
so a client reopening on receipt of exited could hit the open guard while
the entry was still set and have its open dropped. Delete first so the
cleared entry is visible by the time the client sees exited.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Opening a terminal whose session has exited left c.terms[id] set to a
no-op (already-exited path) or to a never-cleared unsubscribe (exit after
open), so the open guard silently dropped every later open for that id on
the connection until close/reconnect. Clients also saw exited/data before
the opened ack.
Ack opened before subscribe so it always precedes replay/data/exited;
have subscribe report whether the pane was already terminal and skip
registering in that case; and clear the connection entry from the exit
callback for panes that exit after open.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The session run loop closes the PTY after copyOut returns, and session.close
(via Manager.Close) closes the same PTY again. creackPTY.Close called cmd.Wait
each time, and a second concurrent Wait on the same process blocks forever, so
daemon shutdown deadlocked whenever a terminal was still attached. fakePTY is
idempotent via sync.Once, so the unit suite never exercised this; a real tmux
attach surfaced it.
Guard close+kill+wait with a sync.Once so Wait runs exactly once. Add a
regression test that double-closes a real PTY under a watchdog.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* feat(backend): HTTP daemon skeleton — config, health, runfile, graceful shutdown (#10)
Phase 1a of the Go HTTP daemon lane (#10). Stands up the loopback-only
sidecar skeleton the later REST/SSE/WS/static surfaces build on:
- config: env-driven (AO_HOST/PORT/ENV/timeouts/run-file) with zero-config
defaults; binds 127.0.0.1:3001; validates and fails fast on bad input.
- httpd: chi router with the recoverer → request-id → logger → real-ip
middleware stack and /healthz + /readyz probes. Per-request timeout is
carried in config but intentionally not global — it scopes to /api/v1 in
Phase 1b so it never throttles SSE/WS/health.
- runfile: atomic PID + port handshake (running.json) for the Electron
supervisor, with a dead-PID stale check so a crashed predecessor doesn't
block startup while a live one fails fast.
- server: bind-before-publish (port conflict fails fast), graceful shutdown
on SIGINT/SIGTERM via signal.NotifyContext with a 10s hard timeout, and
run-file cleanup on exit.
Why: the daemon must be safely supervisable as a child process — the
supervisor needs a discoverable PID/port and the daemon must not leave a
half-started process or stale handshake behind. Locking the lifecycle down
now keeps the future port split a small change rather than a rewrite.
Tests cover config defaults/overrides/validation, run-file round-trip and
live/dead PID detection, health probes, full Run lifecycle, and port-conflict
fail-fast.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* refactor(backend): drop Env config field — not needed yet (#10)
Per review on #14: AO_ENV / Config.Env / IsProduction() weren't load-bearing
for Phase 1a — they only switched the slog handler. Removing them now keeps
the surface minimal; the env knob can come back later when a real consumer
needs it.
- config: remove Env field, AO_ENV parsing, and IsProduction helper.
- main: collapse newLogger to a single text-handler path.
- httpd: drop the env field from the listening log line.
- tests: drop the env assertions and AO_ENV fixture.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* docs: add backend run + config quick-start to README (#10)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(backend): address Phase 1a review comments (#10)
- config: drop AO_HOST entirely — the daemon is loopback-only by design,
so making the bind host env-configurable was a security footgun
- config: use net.JoinHostPort in Addr() so IPv6 literals stay valid
- config: reject zero/negative AO_REQUEST_TIMEOUT and AO_SHUTDOWN_TIMEOUT
(time.ParseDuration accepts both; either would silently break the
daemon — instant request expiry / no graceful drain)
- runfile: split processAlive into unix/windows build-tagged files so
liveness detection is reliable on both platforms (Windows uses
OpenProcess; POSIX keeps signal 0)
- runfile: document os.Rename overwrite semantics (atomic on POSIX,
REPLACE_EXISTING on Windows) so the temp-then-rename pattern's
cross-platform behaviour is explicit
- httpd tests: give probe/waitForHealth clients an explicit per-request
timeout so a stalled connect can't hang the test on the outer deadline
* fix(backend): strip trailing blank line from runfile.go (#10)
gofmt CI was failing because removing the orphan processAlive doc
comment left an extra newline at EOF.
* fix(backend): cross-platform run-file replace + AO_HOST rationale (#10)
- runfile: introduce build-tagged atomicReplace — POSIX rename(2) on
Unix, MoveFileEx with MOVEFILE_REPLACE_EXISTING on Windows. The Go
runtime happens to do the Windows call internally already, but
invoking it directly makes the cross-platform contract explicit
instead of a runtime implementation detail
- runfile: tighten process_unix.go build tag from `!windows` to `unix`
so plan9/js/wasm fail to build rather than silently using a broken
signal-0 probe
- runfile: add TestWriteOverwritesExisting covering the stale run-file
replace path that none of the previous tests exercised
- config: anchor the loopback-only decision in the LoopbackHost doc so
the next contributor doesn't reintroduce AO_HOST without the security
rationale
* fix(backend): route chi access logs through slog/stderr (#10)
chi's middleware.Logger writes via stdlib log to stdout, but the
daemon's slog logger writes to stderr — so REST traffic and daemon
logs landed on different streams in different formats. Replace it
with a small slog-backed requestLogger that:
- Wraps the response writer via middleware.NewWrapResponseWriter so
status/bytes are accurate even when handlers return without an
explicit WriteHeader.
- Reads the request id off the context set by middleware.RequestID
(kept mounted just before this middleware so the id is available).
- Emits one structured Info line per request with method, path,
status, bytes, duration, and remote — same key=value shape as the
rest of the daemon, one stream for the Electron supervisor to
capture.
* feat(api): projects route shell (7 routes, REST-corrected) — #20
Mounts the /api/v1 surface on the skeleton router (#10·1a) and registers
the 7 canonical project routes as 501 stubs that emit a structured
PlannedRoute body documenting the future contract. Shared scaffolding
landed here (api.go, errors.go, stubs/, controllers/) so #21/#22 plug in
without re-touching the wiring.
WHY: opens the route-shell PRs in the Go HTTP daemon lane. Doing it
interface-first lets the dashboard team build against the contract
before any handler logic exists; the locked APIError envelope and
PlannedRoute shape become #19's OpenAPI source-of-truth.
REST audit corrections vs the legacy TS surface:
R3 PUT /projects/:id alias of PATCH: PUT not registered → 405.
R4 POST /projects/:id repair overload: canonical /repair; legacy 405.
R5 degraded GET returns 200 with error field: discriminator status.
R6 ok/success flag flips: drop on 2xx; return affected resource.
R9 bare {error: msg}: locked {error,code,message,requestId,details?}.
Legacy paths are deliberately NOT registered; each canonical handler
carries PlannedRoute.Legacy so consumers can discover the migration.
Zod schemas (TrackerConfig, SCMConfig, AgentConfig, ReactionConfig,
LocalProjectConfig, RoleAgentConfig) ported to typed Go structs with an
Extra map reserved for .passthrough() round-tripping in later PRs.
Closes part of #18; targets feat/issue-10 until #14 merges.
* refactor(api): collapse ProjectService → ProjectManager — #20
Controllers now depend on ONE inbound interface per resource — ports.ProjectManager —
mirroring the existing ports.SessionManager + LifecycleManager pattern.
Whether the manager impl reaches into the registry, the LCM, an outbound
port, or all three is its own concern; the HTTP layer no longer has to
know any of that.
WHY: the original split named the boundary type "ProjectService" and put
it in a sibling services.go. That implied a second category of port
distinct from inbound.go's *Manager interfaces, even though they play
the same role (things HTTP/CLI call into the core). Per review feedback,
collapse them onto one Manager-per-resource pattern.
Mechanical changes:
- ports/inbound.go gains ProjectManager next to SessionManager.
- ports/services.go renamed to projects.go; keeps only the DTOs the
ProjectManager methods take/return.
- ProjectsController.Svc renamed to Mgr; APIDeps.Projects type bumped
to ports.ProjectManager.
All tests pass unchanged; no behavioural change.
* refactor(api): replace stubs/ with OpenAPI-as-source-of-truth — #20
The first cut of the route shell duplicated each route's contract twice:
once as a Go literal (stubs.PlannedRoute{...}) in the controller, and
implicitly in the PR description. The Go literal was ~230 LoC of pure
throwaway that would be deleted in handler-impl PRs.
This commit eliminates the duplication:
- backend/internal/httpd/apispec/openapi.yaml: full OpenAPI 3.1 doc
covering the 7 project routes + shared schemas (Project, APIError,
config types). x-replaces records the legacy → canonical mapping
REST-audit corrections produced.
- apispec/apispec.go: //go:embed the YAML, expose Operation(method,
path) → the spec slice as a map, NotImplemented(w, r, method, path)
→ 501 with that slice embedded as `spec`.
- controllers/projects.go: each of 7 handlers is now a one-liner:
apispec.NotImplemented(w, r, "GET", "/api/v1/projects").
- /api/v1/openapi.yaml serves the embedded document so tooling
(SDK gen, the validator slated for #19, dashboard dev tools) can
fetch the whole spec from the same origin as the routes.
- stubs/ package deleted.
When a real handler lands, only the apispec.NotImplemented line goes
away — nothing else does. The spec stays as documentation; consumers
never had to know it was throwaway. #19 (OpenAPI follow-up) is now
half-folded into this PR; the validation middleware remains its own
follow-up.
Tests reshaped: assert envelope + spec.operationId + spec.x-replaces
(replaces the old planned.legacy assertion); add TestOpenAPIYAMLServed
to cover the static spec serve; add apispec_test.go for embed/lookup
behaviour.
* refactor(api): move projects contract to internal/project package — #20
Pilots the feature-package layout the backend is migrating toward: a
resource's inbound interface and its DTOs live with the resource, not in
a central ports/ catch-all.
WHY: review flagged ports/ as vague. It conflates three jobs — the
outbound capability seam (legit), single-impl inbound interfaces (Go
idiom wants these consumer-side), and DTOs that aren't ports at all.
This moves the projects contract out as the reference shape #21/#22
follow; the merged session/lifecycle/outbound contracts are left
untouched and migrated separately.
Scope: INTERFACE ONLY. No implementation — handlers still answer via
apispec.NotImplemented and the injected project.Manager stays nil. The
impl lands in a later handler-impl PR.
Changes:
- new internal/project: project.go (Manager interface, 7 endpoints) +
dto.go (AddInput/GetResult/UpdateConfigInput/RemoveResult/ReloadResult,
moved verbatim from ports/projects.go, Project-prefix dropped).
- ports/projects.go deleted; ProjectManager removed from ports/inbound.go.
outbound.go and facts.go untouched.
- controllers/projects.go and httpd/api.go depend on project.Manager.
Domain entities (Project, ProjectSummary, DegradedProject, config types)
stay in domain/ as shared vocabulary.
go build/vet/test/gofmt all clean; no behavioural change.
* refactor(api): consolidate project types into internal/project — #20
Addresses PR review: (1) "why are config_types required at the moment?"
and (2) "project objects already defined in project/ — how do we
differentiate?"
Both had the same root cause: project types were split across domain/
and project/. Fix — keep ALL project types in the project package; only
domain.ProjectID (shared with sessions/lifecycle/workspace) stays in
domain.
- domain/project.go → project/types.go: Project, Summary, Degraded
(renamed from ProjectSummary/DegradedProject; the package name carries
the "Project" prefix now).
- domain/config_types.go deleted. Kept only the 4 shapes the projects
API actually exposes — TrackerConfig, SCMConfig, SCMWebhookConfig,
ReactionConfig — moved into project/types.go. Dropped AgentConfig,
AgentPermission, RoleAgentConfig, LocalProjectConfig (zero references)
and the speculative `Extra map[string]any` passthrough fields (no
marshaller existed, so they silently dropped data — premature).
- project/dto.go + project/project.go reference the local types; ids
stay domain.ProjectID.
Net: one home for project types, no dead code. go build/vet/test/gofmt
clean; no behavioural change (handlers still 501 via apispec).
* feat(api): implement project routes with mock manager/store
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* merge: resolve conflicts with origin/main
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* refactor(httpd): share JSON/API error envelope helpers
* fix(api): align project mock store with sqlite schema
* fix(api): address project API review semantics
* canonicalize both paths with filepath.EvalSymlinks before comparing
* style(project): gofmt git repo validation
---------
Co-authored-by: Aditi Chauhan <aditi1178@gmail.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Vaibhaav <user@example.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
If startSession returned an error, run() returned immediately and the
reaper + cdc poller goroutines kept running while defer store.Close()
fired — a data race against the SQLite handle. Mirror the bottom-of-run
shutdown sequence on the error path (cancel ctx, drain reaper, drain
poller) so both goroutines have exited before the store is closed. The
explicit-not-defer ordering is the same the existing
post-srv.Run shutdown block uses; piling on more defers would hit the
LIFO trap the same comment already warns about.
Reported by Greptile on PR #52.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Renames the unused context.Context parameter from `_` to `ctx` so the
parameter name is already in place when a future plugin constructor
needs to honor cancellation (tmux/gitworktree are synchronous today).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Constructs a live *session.Manager in main alongside the LCM, sharing the
exact same SessionStore + LCM dependencies the lifecycle stack already
holds.
Refactor: storeAdapter moves from package main to a new internal
package, wiring.Adapter, so the daemon's composition root and any
in-process integration tests can share a single bridge.
Stubbed for now: ports.Agent has no production adapter on main; a loud
*noopAgent returns sentinel AO_AGENT_HARNESS_NOT_WIRED and logs a
warning once on first call, so a future Spawn through this lane fails
at the runtime layer with a clear breadcrumb rather than starting a
broken session quietly. ports.Notifier and ports.AgentMessenger remain
stubbed alongside the LCM.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
A new subscriber could receive the same PTY bytes twice: the ring snapshot
was taken under s.mu but replayed after unlock, while copyOut appended to the
ring and fanned out as two separate lock acquisitions. A chunk appended before
the snapshot could then be fanned out after the replay, delivering it in both.
The same gap let an exited frame overtake the replay.
Make append+fanout one atomic step under s.mu (deliver) and replay the snapshot
before releasing the lock in subscribe, so the two critical sections fully
serialize and each chunk reaches a subscriber exactly once.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Construct the tmux runtime and a terminal.Manager fed by the CDC
broadcaster, and hand it to httpd.New so the /mux WebSocket surface goes
live. httpd.New now runs after the CDC substrate so the broadcaster exists
when the manager subscribes; the listener still binds before running.json
is written, preserving fail-fast on port conflict. The manager is closed on
shutdown alongside the CDC pipeline and lifecycle stack.
Add the /mux route: httpd performs the WebSocket upgrade (coder/websocket)
and adapts the connection to terminal.wsConn via wsjson, then hands it to
terminal.Manager.Serve. httpd owns only the upgrade and transport
adaptation; all stream logic stays in internal/terminal.
The route is mounted outside the per-request Timeout middleware (the
connection is long-lived) and is omitted entirely when no manager is wired,
so the daemon degrades to no terminal surface rather than failing. New/
NewRouter take the manager; main.go passes nil until commit 3 wires it.
mux_test.go drives the real upgrade + wsjson + Serve + creack/pty path with
a throwaway shell command, so it needs no tmux.
Add internal/terminal: a transport-agnostic feature package that attaches
a PTY to a session's tmux pane and multiplexes the byte stream to WebSocket
clients, plus a session-state channel fed by the CDC broadcaster.
The PTY is reached through a small PTYSource interface (satisfied by the
tmux runtime adapter) and spawned via an injectable spawnFunc, so fan-out,
the 50KB replay ring, and re-attach resilience all test without a real
process, tmux, or network. A tmux-guarded integration test exercises the
real creack/pty path end-to-end.
Raw PTY bytes never touch the CDC change_log; only the sessions channel is
CDC-fed. Windows PTY spawning is stubbed pending a ConPTY path.
Greptile P1: the patch-then-Update path in TestRestoreRoundTrip was
discarding GetSession's ok/err. A missed row would have handed UpdateSession
a zero-value SessionRecord (ID==""), which matches zero rows and returns
nil — Phase B then fails with the misleading "agent session id lost across
restart" instead of the real cause.
Fixed at the patch site (the only write path that could swallow the error)
and at the two read-then-assert sites in TestDetectingPersistsAcrossRestart
for consistency. Downstream assertions there would already fail loudly, but
the explicit ok/err check makes the failure mode unambiguous.
All 219 tests still pass under -race.
- Idempotent close + t.Cleanup so a mid-test t.Fatalf can't leak the
SQLite handle for the rest of the binary run. Restart-style tests still
call close() explicitly between phases; the cleanup hook becomes a no-op
once that runs.
- Drop the hardcoded "mer-1" assertion in TestHappyPath; assert the
structural invariant (project-scoped, non-empty id) instead, so the
test does not couple to the {project}-{counter} generation detail.
- Realign the test storeAdapter's PRFactsForSession + WritePR bodies to
be line-for-line identical to backend/lifecycle_wiring.go's production
adapter (extracted prState helper, separated err vs empty-rows checks),
so a future divergence shows up as a diff at review time. The proper
fix (extract to internal/storeutil) remains out of scope per the
brief's "do NOT redesign anything".
All 219 tests still pass under -race.
The Check formatting step failed on the stub method alignment in
stubAgent's three method declarations. Re-run of gofmt aligns the column
gutter on those signatures; no behavioural change.
Adds backend/internal/integration with five end-to-end tests that hydrate the
real lifecycle.Manager + session.Manager against a tmp SQLite store and
exercise the full pipeline through the DB triggers and the CDC poller:
- TestHappyPath_Spawn_PR_Kill — spawn -> SCM PR observation (open + CI
passing) -> kill; asserts canonical row, pr row, and change_log event
types (session_created/_updated, pr_created, pr_check_recorded).
- TestRestoreRoundTrip_PreservesMetadata — spawn, kill, close store, reopen
same DB path, hydrate fresh LCM/SM, Restore(); asserts AgentSessionID and
the rest of SessionMetadata survive across the daemon restart.
- TestCIFailureAndRecovery_NudgeThenClears — failing CI observation drives
the CI-failed reaction nudge with the log tail injected; passing CI
observation switches to approved-and-green human notify; pr_checks history
reads back the failure (the brake's source of truth).
- TestDetectingPersistsAcrossRestart — failed probe parks the session in
detecting with detecting_* columns populated, round-trips across a
close/reopen, alive probe clears the quarantine memory.
- TestCDCPollerReceivesAllStages — drives the real cdc.Poller; asserts the
trigger pipeline emits each expected event_type and seq is monotonic.
Wiring gap fixed (minimal): goose v3 keeps baseFS/logger/dialect as
package-level globals, so two concurrent sqlite.Open() calls — uncommon in
production but normal under -race with t.Parallel() — race on
goose.SetBaseFS/SetLogger/SetDialect inside migrate(). Added a process-level
sync.Mutex around the migrate() call. ~11 lines, no signature changes.
Scope notes (the task brief assumed a fancier architecture than what
actually shipped in PR #37):
- No outbox / consumer_offsets / janitor exist on main — the change_log
table IS the durable, ordered source of truth (see cdc/event.go), so the
brief's janitor-watermark step is skipped.
- No reaction_trackers table / ReactionStore port — trackers are in-memory
per lifecycle/reactions.go; persistence-round-trip there is N/A.
- No revision column / Upsert(rec, eventType) — write-mutex serialises and
change_log.seq orders, so the assertions land on event_type + seq, not on
a per-row revision counter.
All 219 tests pass under -race across 18 packages. lifecycle/fakes_test.go
is untouched; existing unit tests still drive the in-memory fake.
Addresses review on PR-observation persistence:
- pr_checks now has an AFTER UPDATE CDC trigger (guarded on status change), so a
check flipping in_progress->failed on the same commit emits change_log instead
of updating silently. Restores symmetry with the sessions/pr triggers.
- writePR persists scalar facts + checks + comments in ONE transaction via
Store.WritePRObservation, so a mid-write failure can't leave the pr row (and
its CDC event) committed while checks/comments are partial. Collapses the
PRWriter port's three write methods into one WritePR.
- db.go: record why modernc.org/sqlite (pure-Go, CGO-free static binary) at the
import site.
Regression tests for both the update-trigger (emit on change, suppress no-op
re-poll) and the transactional write. go test -race ./... green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
lcStack.Stop()/cdcPipe.Stop() block on done channels that close only when ctx
is cancelled, but the deferred stop() that cancels ctx ran last (LIFO) — so any
non-signal exit (e.g. a listener bind error) hung the daemon forever. Cancel ctx
first, then drain, explicitly after srv.Run instead of via defer. Also refresh
the startup comments that still described the removed outbox/JSONL/janitor flow.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reworks the LCM, reactions, session manager, reaper, and boot wiring onto
the redesigned domain model — collapsing the runtime axis to is_alive,
moving PR facts to the pr table (read back as PRFacts), replacing the
free-form SessionReason with a typed terminal-only TerminationReason, and
dropping Revision/EventType/durable reaction-trackers (CDC is trigger-driven,
escalation budgets are in-memory).
- ports: SessionStore + PRWriter interfaces; PRObservation/RuntimeFacts/
ActivitySignal DTOs; drop LifecycleStore/EventType/ReactionStore.
- lifecycle: single-writer reducer over is_alive; ApplyPRObservation writes
the pr tables and reacts; CI-fix-loop brake derived from pr_checks history;
review comments injected into the agent regardless of author (no bot
detection); merge auto-terminates with pr_merged.
- session: store-assigned "{project}-{n}" ids; folded metadata; status
derived from PRFacts on read.
- reaper: reports the four-valued probe vocabulary unchanged.
- boot: trigger -> poller -> broadcaster; storeAdapter bridges *sqlite.Store.
Lane shrinks 6218 -> 2803 LOC. go build/vet/test -race green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reworks the storage + CDC layer to the simplified design agreed in review:
Schema (one clean migration, 0001): projects, sessions, pr, pr_checks,
pr_comment, change_log. sessions.id is a single string key "{project}-{num}"
(mer-1); operational metadata folded into sessions; is_alive replaces the
runtime axis; no revision (the per-session write mutex serializes, change_log.seq
orders). pr keyed by URL (1 session : many PRs). pr_checks is CI run history
(one row per check per commit) — the CI-fix-loop brake is a LIMIT 3 query, no
counter stored. change_log carries a required project_id FK + nullable session_id.
CDC is DB-native: AFTER INSERT/UPDATE triggers on sessions/pr/pr_checks append
to change_log atomically with the change (json_object payloads). The old durable
outbox/JSONL/janitor pipeline is gone; the cdc package is now a Poller that reads
change_log and fans events out through the in-memory Broadcaster (hardened with
recover()). Clients catch up via the log from their own offset (SSE Last-Event-ID).
Storage uses a single writer connection + a reader pool (read-your-writes for the
triggers' subqueries; concurrent reads). sqlc-generated typed queries.
Tests (-race): CRUD, per-project id assignment, the loop-brake query, concurrent
creates, triggers populating change_log; CDC end-to-end through the real store,
concurrent goroutine delivery, broadcaster panic-isolation.
NOTE: scoped to storage + CDC. The lifecycle-engine consumers (decide, lifecycle,
session, reaper, main wiring) still reference the old domain axes and need a
follow-up integration pass to compile against the new model.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
SetMaxOpenConns(1) forced every read (List/Get/GetPR/...) to queue behind
the single connection, so the dashboard's reads contended with the LCM's
writes. WAL already supports many concurrent readers, so raise the pool to 8
and instead serialize *writes* with a Store.writeMu. That keeps WAL's
single-writer rule and the revision-CAS read-then-write atomic regardless of
pool size, while reads now run in parallel across the pool.
Every write method takes writeMu (Upsert, PatchMetadata, UpsertPR/DeletePR,
the pr_check/pr_comment Replace* via inTx, the CDC outbox/offset writes,
project writes, reaction-tracker writes); reads take nothing. Added
TestConcurrentReadsAndWrites (16 writers + 16 readers) which passes under
-race.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The cdc integration test covers the synchronous Drain/Poll happy path but
(1) resyncs from a fake snapshot and (2) never runs the publisher/consumer as
the concurrent goroutines the daemon actually uses. Add two E2E tests in the
composition-root package that wire the real sqlite.Store, outboxAdapter,
Publisher, JSONL log, Consumer, Broadcaster and the REAL snapshotSource
(store.ListAll):
- RealSnapshotResyncThroughRotation: forces a rotation and asserts the
consumer rebuilds from the sessions table, delivering the persisted record
payload, with the offset landing at the change_log head.
- ConcurrentPublisherConsumer: runs both as goroutines on their tickers and
asserts every write is delivered exactly once, in order, offset at head
(also exercises the broadcaster hand-off under -race).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The first storage cut modelled two side tables as free-form blobs. This
replaces both with opinionated, statically-typed schema so what a session
can carry is fixed by the schema, not by convention.
session_metadata: was a (session_id, key, value) KV bag with six
convention-only keys. Now a 1:1 table of named, typed columns. The domain
currency is a typed domain.SessionMetadata struct (was map[string]string),
threaded through ports.LifecycleStore, the LCM, the Session Manager and the
reaper, so an unknown key is a compile error rather than a silently-dropped
write. PatchMetadata keeps its non-destructive merge ("empty = leave
unchanged"). The off-canonical invariant is now enforced at the type level
via json:"-" on SessionRecord.Metadata, removing the manual `Metadata = nil`
scrub the change_log/snapshot paths had to remember; the Meta* string-key
constants are deleted.
pr_enrichment -> pr (+ pr_check, pr_comment): the scalar facts are now
typed columns with CHECK-constrained enums (review_decision, mergeability,
ci_state) and integer CI counts instead of opaque TEXT. The two list facts
the old `pending_comments`/ci_summary strings smuggled are normalized into
child tables (pr_check, pr_comment) that cascade from pr. The store exposes
UpsertPR/GetPR plus atomic ReplacePRChecks/ReplacePRComments + List.
Both tables remain off the canonical CDC path. sqlc regenerated; migrations
0001/0002 revised in place (nothing released). gofmt/vet clean; go test
-race green; daemon smoke-boots and creates the new schema.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Six fixes from the second code review pass — none load-bearing but all
improve the contract honesty or prevent future churn. Tests pass with
-race (49 in package, 299 across the backend).
1. Preflight: atomic.Bool fast-path before the mutex so cached-success
calls don't contend on the lock. Double-checked locking on the
mutex side so concurrent first-callers still serialize the single
GET /user request.
2. Preflight godoc: tighten to say it verifies token validity, not
repo authorization — Get/List against a specific repo may still
return ErrAuthFailed after a green Preflight if the token lacks
the scope or the repo isn't visible.
3. domain.ListFilter.Limit godoc: explicitly note that exceeding the
provider per-page max is SILENTLY capped (no error, no truncation
signal) and that auto-pagination is deferred to #35.
4. Extract issueFromGH helper. Get and List were duplicating the
identical ghIssue -> domain.Issue projection; consolidating now
prevents a 3-way merge when #40 (Comment/Transition) lands.
5. parseGitHubRepo: reject whitespace and # in both owner and repo
segments. Leading dots stay legal (the "owner/.github" repo
convention). New test cases cover the rejections plus a positive
guard for the leading-dot case.
6. fakeGH test helper: lock the handlers map on both read and write,
so future tests registering handlers from goroutines won't trip
-race.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Brings the read-side surface up to parity with the legacy TS impl's
useful read methods. Closes a gap flagged during scope review.
Port additions:
- List(ctx, repo, filter) ([]Issue, error)
- Preflight(ctx) error
Domain additions: TrackerRepo, ListStateFilter (open/closed/""=all),
ListFilter (State, Labels, Assignee, Limit).
GitHub adapter:
- List hits GET /repos/{o}/{r}/issues with query-encoded filter.
Defaults: state=all, per_page=30; per_page is capped at 100.
PRs are filtered out client-side (GitHub conflates them with
issues on that endpoint).
- Preflight hits GET /user. Success is cached for the Tracker's
lifetime via sync.Mutex + bool; failures are intentionally NOT
cached so a transient startup glitch is recoverable.
- New ErrAuthFailed sentinel. classifyError now maps 401, and 403
without rate-limit signals, to ErrAuthFailed instead of an opaque
error — so Preflight callers can distinguish bad-token from other
failures.
Pagination beyond the first page is intentionally out of scope for v1
(see doc.go); the observer/polling work in #35 will own that.
Tests: 43 pass (was 25). Adds Preflight cache + recovery, List query
encoding, PR filtering, repo parser rejection, and ErrAuthFailed
classification.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Scope correction: mirroring agent lifecycle onto the tracker (status
comments, label/state updates) is not wanted in the current rewrite.
That work is now tracked as issue #40 and will land once we decide on
the opt-out knob, label setup, and Linear's workflow-state fit.
Removes from the port and the GitHub adapter:
- Comment(ctx, id, body) and ErrEmptyBody
- Transition(ctx, id, state) and ErrUnknownState
- planForState / transitionPlan and the forward state mapping
- reasonComplete constant (only the Get reverse mapping is kept)
- 11 tests + the transitionCall normalization helpers
Kept (still load-bearing for Get):
- All 5 NormalizedIssueState values — Get reports them faithfully
when a repo carries the in-progress / in-review labels.
- The reverse mapping in mapStateFromGitHub.
- RateLimitError with ResetAt + RetryAfter (#35 will use it).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Migration 0002 adds two tables off the canonical CDC path:
- projects: durable registry of managed repos (the twin of the old YAML
config). Soft-deletable via archived_at so a session's project_id always
resolves; ListProjects returns active rows only, GetProject resolves any.
- pr_enrichment: per-session cache of rich SCM facts (CI summary, review
decision, mergeability, pending comments, CI log tail) that do not live
in the canonical lifecycle. 1:1 with a session, cascades on session delete.
Both are written outside the LCM write path: no revision bump, no
change_log/outbox event. Store methods mirror the reaction_trackers adapter
pattern with storage-local row structs.
Add the two real outbound adapters that replace the in-memory fakeStore:
internal/storage/sqlite (persistence satisfying ports.LifecycleStore) and
internal/cdc (transactional-outbox publisher, JSONL delivery, durable
consumer). Wire them into main.go alongside the Lifecycle Manager and reaper
so the write path is live end-to-end: LCM.Upsert -> store -> outbox -> JSONL
-> broadcaster.
Storage (internal/storage/sqlite):
- modernc.org/sqlite (pure Go, no CGO) for clean cross-compile; goose
embedded migrations; sqlc-generated typed queries under gen/.
- Atomic Upsert: session row + change_log + outbox written in one tx.
- revision is an optimistic-concurrency (CAS) check: insert requires
revision 0 and persists 1; update requires loaded revision == stored and
bumps +1; zero rows affected returns a revision-mismatch error.
- Metadata is an opaque map in session_metadata, off the CDC path.
- Durable reaction_trackers (fixes the in-memory-only escalation budget that
re-fired human pages on restart).
CDC (internal/cdc):
- Publisher drains the outbox to a JSONL log; size-based rotation with a
reset marker.
- Consumer tails via byte cursor, detects rotation (os.SameFile), resyncs
from a full-state snapshot on gaps, and tracks a durable consumer_offsets
cursor.
- Janitor reclaims acknowledged outbox rows.
- Broadcaster is the in-process fan-out port the FE transport will subscribe
to (WS/SSE wiring deferred).
Composition root (main.go + *_wiring.go):
- startCDC stands up publisher/consumer/janitor + broadcaster.
- startLifecycle constructs the LCM, makes escalation budgets durable via
WithReactionStore, teaches it to enumerate sessions via WithSessionLister,
and starts the reaper.
- Notifier, AgentMessenger, and the reaper's runtime registry are TEMPORARY
no-op/empty stubs (lifecycle_wiring.go) with TODO markers; see the PR
description for how to fill them in.
Tests: contract-parity, revision CAS, outbox atomicity, CDC ordering and
idempotency, rotation/resync, janitor vacuum, reaction durability across a
simulated restart, and composition-root adapters. gofmt/build/vet clean and
go test -race ./... green.
Reference implementation; GitLab and Linear follow in separate PRs.
Issue observer loop (poll + ApplyTrackerFacts) is deferred to #35.
Three-layer split mirrors the SCM layout adil is landing in PR #28:
- domain/tracker.go — value types (TrackerProvider, TrackerID,
NormalizedIssueState, Issue)
- ports/tracker.go — the Tracker interface
- adapters/tracker/github/ — REST-backed adapter
v1 is write-mostly: Get, Comment, Transition. No cache, no inflight
dedup, no polling. State mapping is documented in the package doc and
exercised by table-driven tests against an httptest fake — no real
GitHub traffic from CI.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Address blocker found in self-review (B1 + I1):
- Manager.RunningSessions previously filtered to runtime.State == RuntimeAlive,
but a session enters Detecting with runtime axis = RuntimeProbeFailed (failed
probe path, decide_bridge.go:72) or RuntimeMissing (detectingLC in
manager_test.go:539). The filter silently parked every Detecting session, so
the recovery path proved by manager_test.go:59 ("healthy probe recovers
liveness-owned detecting -> working") and the terminal path proved by
manager_test.go:79 ("dead+dead with no recent activity concludes killed")
were both unreachable through the reaper. Broaden the predicate to "session
is not in a terminal state" (mirrors the LCM's existing isTerminal helper)
and document the wider semantics.
- reaper.probeOne now reports every probe result — including alive — back to
the LCM as ApplyRuntimeObservation facts. The previous skip-alive
optimization was a layering violation: the reaper has no business deciding
what counts as a no-op. The LCM's ApplyRuntimeObservation already diffs
against canonical and only Upserts on actual change, so steady-state alive
stays cheap. With the broadened poll set, an alive probe for a Detecting
session IS the recovery fact.
- Add unit tests for Manager.RunningSessions covering: nil-lister no-op, lister
error propagation, and the full canonical state matrix (working/idle/
needs_input/detecting-probefailed/detecting-missing/not_started included;
terminated/done excluded).
- Update reaper tests: alive case now asserts the alive fact is reported; new
"detecting session: alive probe reported so LCM can recover from quarantine"
case locks in the recovery path; multi-runtime case now asserts both runtime
facts flow through.
- Bump "session in poll set without handle metadata" log from Debug to Warn —
it is an anomaly (OnSpawnCompleted should have written both keys), not a
routine event.
- Document WithSessionLister must be called before any reaper attached to the
Manager starts running (it is a bare field read; concurrent re-injection is
meaningless anyway).
The reaper sits OUTSIDE the LCM's per-session serial loop. On every tick it:
1. Fires lcm.TickEscalations(now) — the duration-based escalation heartbeat
a non-polling LCM cannot wake itself to drive.
2. Asks lcm.RunningSessions for the snapshot of sessions whose runtime axis is
alive, then calls runtime.IsAlive(handle) per session via a RuntimeRegistry
that dispatches by RuntimeHandle.RuntimeName (so a single reaper covers
tmux + zellij side by side).
3. Reports any non-alive result back as a fact via ApplyRuntimeObservation —
dead -> RuntimeProbeDead, probe error -> RuntimeProbeFailed (never
collapsed to alive: failed probe ≠ dead, but it ≠ alive either). Steady-
state alive is skipped so we don't churn the LCM with no-op load/diff work.
The reaper REPORTS facts; the LCM owns DECIDE (anti-flap Detecting quarantine,
terminal-session rules). The reaper never writes.
Open-question resolution: add RunningSessions(ctx) to ports.LifecycleManager
(option a). The Manager implements it via an injectable session lister
(Manager.WithSessionLister) so the LCM itself does not require a new
LifecycleStore method — Tom's store contract is untouched, daemon wiring (lane
#10) will inject the production lister at startup.
Scope: reaper goroutine + the minimum LCM seam. No activity ingest, no FS
watcher, no daemon wiring, no new schema fields, no store changes.
Four narrowly-scoped fixes against the LCM + Session Manager lane
from an external review of the current backend state. R2 (failed-restore
lifecycle stranding) is intentionally deferred to PR #15, which already
closes it via the new OnSpawnInitiated path; R3 also stays on that PR.
- R1 (BLOCKER): Manager.Spawn never persisted AgentSessionID, so
Manager.Restore's hard-required metadata key was always missing and
every restore failed. Persist the assembled launch prompt as
MetaPrompt at spawn time and add a fresh-launch fallback to Restore
that uses Agent.GetLaunchCommand with the seeded prompt when the
captured agent session id is absent (the id-capture hook is a separate
path that may never have run). Restore still fails fast when neither
the id nor a prompt is on hand — there is nothing to relaunch from.
- RA (BLOCKER): adapters/workspace/gitworktree/commands.go's
worktreeRemoveForceArgs passed --force, which deletes uncommitted
agent work. Renamed to worktreeRemoveArgs and dropped --force so the
post-prune "still registered" guard in Workspace.Destroy surfaces the
refusal to Manager.Cleanup, which routes the session to Skipped
instead of destroying in-progress changes.
- R11 (SHOULD-FIX): reactions.go's two Notifier.Notify call sites
(executeReaction's notify and escalate) built OrchestratorEvent
without ProjectID. Captured projectID on the transition (via a
store.Get in mutate) and on reactionTracker (so TickEscalations can
still populate it on duration-based escalations), and threaded it
through executeReaction/sendToAgent/escalate.
- RB (SHOULD-FIX): gitworktree.Workspace.managedPath used filepath.Join
which cleans .. segments before validateManagedPath ran, so
session=\"../other\" stayed inside managedRoot while breaking
per-project isolation. validateConfig now rejects path separators and
the . / .. components on ProjectID and SessionID at the source.
go build ./..., go vet ./..., and go test -race ./... all pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(backend): HTTP daemon skeleton — config, health, runfile, graceful shutdown (#10)
Phase 1a of the Go HTTP daemon lane (#10). Stands up the loopback-only
sidecar skeleton the later REST/SSE/WS/static surfaces build on:
- config: env-driven (AO_HOST/PORT/ENV/timeouts/run-file) with zero-config
defaults; binds 127.0.0.1:3001; validates and fails fast on bad input.
- httpd: chi router with the recoverer → request-id → logger → real-ip
middleware stack and /healthz + /readyz probes. Per-request timeout is
carried in config but intentionally not global — it scopes to /api/v1 in
Phase 1b so it never throttles SSE/WS/health.
- runfile: atomic PID + port handshake (running.json) for the Electron
supervisor, with a dead-PID stale check so a crashed predecessor doesn't
block startup while a live one fails fast.
- server: bind-before-publish (port conflict fails fast), graceful shutdown
on SIGINT/SIGTERM via signal.NotifyContext with a 10s hard timeout, and
run-file cleanup on exit.
Why: the daemon must be safely supervisable as a child process — the
supervisor needs a discoverable PID/port and the daemon must not leave a
half-started process or stale handshake behind. Locking the lifecycle down
now keeps the future port split a small change rather than a rewrite.
Tests cover config defaults/overrides/validation, run-file round-trip and
live/dead PID detection, health probes, full Run lifecycle, and port-conflict
fail-fast.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* refactor(backend): drop Env config field — not needed yet (#10)
Per review on #14: AO_ENV / Config.Env / IsProduction() weren't load-bearing
for Phase 1a — they only switched the slog handler. Removing them now keeps
the surface minimal; the env knob can come back later when a real consumer
needs it.
- config: remove Env field, AO_ENV parsing, and IsProduction helper.
- main: collapse newLogger to a single text-handler path.
- httpd: drop the env field from the listening log line.
- tests: drop the env assertions and AO_ENV fixture.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* docs: add backend run + config quick-start to README (#10)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(backend): address Phase 1a review comments (#10)
- config: drop AO_HOST entirely — the daemon is loopback-only by design,
so making the bind host env-configurable was a security footgun
- config: use net.JoinHostPort in Addr() so IPv6 literals stay valid
- config: reject zero/negative AO_REQUEST_TIMEOUT and AO_SHUTDOWN_TIMEOUT
(time.ParseDuration accepts both; either would silently break the
daemon — instant request expiry / no graceful drain)
- runfile: split processAlive into unix/windows build-tagged files so
liveness detection is reliable on both platforms (Windows uses
OpenProcess; POSIX keeps signal 0)
- runfile: document os.Rename overwrite semantics (atomic on POSIX,
REPLACE_EXISTING on Windows) so the temp-then-rename pattern's
cross-platform behaviour is explicit
- httpd tests: give probe/waitForHealth clients an explicit per-request
timeout so a stalled connect can't hang the test on the outer deadline
* fix(backend): strip trailing blank line from runfile.go (#10)
gofmt CI was failing because removing the orphan processAlive doc
comment left an extra newline at EOF.
* fix(backend): cross-platform run-file replace + AO_HOST rationale (#10)
- runfile: introduce build-tagged atomicReplace — POSIX rename(2) on
Unix, MoveFileEx with MOVEFILE_REPLACE_EXISTING on Windows. The Go
runtime happens to do the Windows call internally already, but
invoking it directly makes the cross-platform contract explicit
instead of a runtime implementation detail
- runfile: tighten process_unix.go build tag from `!windows` to `unix`
so plan9/js/wasm fail to build rather than silently using a broken
signal-0 probe
- runfile: add TestWriteOverwritesExisting covering the stale run-file
replace path that none of the previous tests exercised
- config: anchor the loopback-only decision in the LoopbackHost doc so
the next contributor doesn't reintroduce AO_HOST without the security
rationale
* fix(backend): route chi access logs through slog/stderr (#10)
chi's middleware.Logger writes via stdlib log to stdout, but the
daemon's slog logger writes to stderr — so REST traffic and daemon
logs landed on different streams in different formats. Replace it
with a small slog-backed requestLogger that:
- Wraps the response writer via middleware.NewWrapResponseWriter so
status/bytes are accurate even when handlers return without an
explicit WriteHeader.
- Reads the request id off the context set by middleware.RequestID
(kept mounted just before this middleware so the id is available).
- Emits one structured Info line per request with method, path,
status, bytes, duration, and remote — same key=value shape as the
rest of the daemon, one stream for the Electron supervisor to
capture.
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Dhruv Sharma
Adil Shaikh
whoisasx
Harshit Singh Bhandari
Pritom Mazumdar
Vaibhaav Tiwari
harshitsinghbhandari
prateek
neversettle