99ff296f6f7aa3e22c738654f6c89b7ce524228f
2 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Alex Newman
|
9e2973059a |
UX redesign: installer + provider rename + /learn-codebase + welcome card + SessionStart hint (#2255)
* feat(ux): claude-mem UX improvements with installer enhancements
Squashed PR #2156 commits for clean rebase onto main:
- feat(installer): add provider selection, model prompt, worker auto-start
- refactor: rename *Agent provider classes to *Provider
- feat: add /learn-codebase skill and viewer welcome card
- feat(worker): inject welcome hint when project has zero observations
- fix(pr-2156): address greptile review comments
- fix(pr-2156): address coderabbit review comments
- fix(pr-2156): persist CLAUDE_MEM_PROVIDER for non-claude in non-TTY mode
- fix(pr-2156): file-backed settings reads in installer + env-first SKILL doc
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* build: rebuild plugin artifacts after rebase onto v12.4.7
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(skills): strip claude-mem internals from learn-codebase
The learn-codebase skill, install next-step copy, WelcomeCard, and
welcome-hint previously walked the primary agent through worker endpoints
and synthetic observation payloads. The PostToolUse hook already captures
every Read/Edit the agent makes — the agent should have no awareness that
the memory layer exists. Collapse the skill to one instruction ("read every
source file in full") and rephrase touchpoints to describe only what the
user observes (Claude reading files), not what happens behind the scenes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(sync): preflight version mismatch + settings-aware port resolution
Two related fixes for build-and-sync's worker restart step:
1. Read CLAUDE_MEM_WORKER_PORT from ~/.claude-mem/settings.json the same
way the worker does, instead of computing the default port from the
uid alone. Previously, users with a custom port saw a misleading
"Worker not running" message because the restart POST hit the wrong
port and got ECONNREFUSED.
2. Add a preflight check that aborts the sync when the running worker's
reported version does not match the version we are about to build.
Claude Code's plugin loader pins the worker to a specific cache
version per session, so syncing into a newer cache directory has no
effect until the user runs `claude plugin update thedotmack/claude-mem`
to bump the pin. The preflight surfaces this explicitly with the exact
command to run; --force bypasses it for intentional cases.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(learn-codebase): note sed for partial reads of large files
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor: strip comments codebase-wide
Removed prose comments from all tracked source. Preserved directives
(@ts-ignore, eslint-disable, biome-ignore, prettier-ignore, triple-slash
references, webpack magic, shebangs). Deleted two tests that asserted
on comment text rather than runtime behavior.
Net: 401 files, -14,587 / +389 lines, -10.4% bytes.
Verified: typecheck passes, build passes, test count unchanged from
baseline (22 pre-existing fails, all unrelated).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(installer): move runtime setup into npx, eliminate hook dead air
Smart-install ran 3 times during a fresh install — the worst run was silent,
fired by Claude Code's Setup hook after `claude plugin install`, producing
~30s of dead air that looked like the plugin was hung.
This change makes `npx claude-mem install` the single place heavy work
happens, with a visible spinner. Hooks become runtime-only.
- New `src/npx-cli/install/setup-runtime.ts` module: ensureBun, ensureUv,
installPluginDependencies, read/writeInstallMarker, isInstallCurrent.
Marker schema preserved exactly ({version, bun, uv, installedAt}) so
ContextBuilder and BranchManager readers keep working.
- `npx claude-mem install`: ungated copy/register/enable for every IDE,
inserts a "Setting up runtime" task with honest "first install can take
~30s" spinner. The claude-code shell-out to `claude plugin install` is
removed — npx already populated everything Claude reads.
- New `npx claude-mem repair` command for post-`claude plugin update`
recovery, force-reinstalls runtime.
- Setup hook now runs `plugin/scripts/version-check.js` (29ms wall) instead
of smart-install. Mismatch prints "run: npx claude-mem repair" on stderr.
Always exits 0 (non-blocking, per CLAUDE.md exit-code strategy).
- SessionStart loses the smart-install entry; 2 hooks remain (worker start,
context fetch).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(installer): delete smart-install sources, retarget tests
- Delete scripts/smart-install.js + plugin/scripts/smart-install.js (both
are source files kept in sync manually; both must go).
- Delete tests/smart-install.test.ts (covered surface is gone).
- tests/plugin-scripts-line-endings: drop smart-install.js entry.
- tests/infrastructure/plugin-distribution: retarget two assertions at
version-check.js (the new Setup hook script).
- New tests/setup-runtime.test.ts: 9 tests covering marker read/write,
isInstallCurrent semantics. Marker schema invariant verified.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(installer): describe npx-driven setup + version-check Setup hook
Sweep public docs and architecture notes to reflect the new flow:
npx installer does Bun/uv setup with a visible spinner; Setup hook runs
sub-100ms version-check.js; users hit `npx claude-mem repair` after a
`claude plugin update`.
- docs/architecture-overview.md: hook lifecycle table + npx flow paragraph
- docs/public/configuration.mdx: tree + hook config example
- docs/public/development.mdx: build output line
- docs/public/hooks-architecture.mdx: full rewrite of pre-hook section,
timing table, performance table
- docs/public/architecture/{overview,hooks,worker-service}.mdx: tree
comments, JSON config example, Bun requirement section
docs/reports/* untouched (historical incident reports).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(install): mergeSettings writes via USER_SETTINGS_PATH
Greptile P1 (#2156): `settingsFilePath()` only resolved
`process.env.CLAUDE_MEM_DATA_DIR`, while `getSetting()` reads via
`USER_SETTINGS_PATH` which `resolveDataDir()` populates from BOTH the env
var AND a `CLAUDE_MEM_DATA_DIR` entry persisted in
`~/.claude-mem/settings.json`. Result: a user with the data dir saved in
settings.json but not exported in their shell would have provider/model
settings silently written to `~/.claude-mem/settings.json` while
`getSetting()` read from `/custom/path/settings.json` — read/write split.
Drop `settingsFilePath()` and the now-unused `homedir` import; reuse the
already-imported `USER_SETTINGS_PATH` constant.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(cli): parse --provider, --model, --no-auto-start install flags
Greptile P1 (#2156): InstallOptions has fields `provider`, `model`,
`noAutoStart`, but the install case in the npx-cli switch only parsed
`--ide`. The other three flags were silently dropped — `npx claude-mem
install --provider gemini` was a no-op.
Extract a `parseInstallOptions(argv)` helper, share it between the bare
`npx claude-mem` and `npx claude-mem install` paths, and validate
`--provider` against the allowed set. Update help text accordingly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(install): pipe runtime-setup output, always show IDE multiselect
Two issues caught in a docker test of the installer:
1. The bun.sh installer, uv installer, and `bun install` were using
stdio: 'inherit', dumping their stdout/stderr through clack's spinner
region — visible as raw "downloading uv 0.11.8…" / "Checked 58
installs across 38 packages…" text streaming under the spinner. Switch
to stdio: 'pipe' and surface captured stderr only on failure (via a
shared describeExecError() helper that includes stdout when stderr is
empty). Spinner stays clean on the happy path.
2. promptForIDESelection() silently picked claude-code when no IDEs were
detected, never showing the user the multiselect. On a fresh machine
with no IDEs present yet (e.g. our docker test container), the user
never got to choose. Now: always show the full IDE list when
interactive; mark detected ones with [detected] hints and pre-select
them; show a warn line if zero are detected explaining they should pick
what they plan to use. Non-TTY callers still get the silent
claude-code default at the call site (unchanged).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(install): skip marketplace work for claude-code-only, offer to install Claude Code
Two related UX fixes from a docker test:
**Delay between "Saved Claude model=…" and "Plugin files copied OK"**
After dropping the needsManualInstall gate, every install was unconditionally
running `copyPluginToMarketplace` (which copied the entire root node_modules
tree — thousands of files, dozens of seconds) and `runNpmInstallInMarketplace`
(npm install --production) even when only claude-code was selected. Neither
is needed for claude-code: that path uses the plugin cache dir + the
installed_plugins.json + enabledPlugins flag, all of which we already write.
- Drop `node_modules` from `copyPluginToMarketplace`'s allowed-entries list;
the dependency-install task populates it on the destination side anyway.
- Re-introduce `needsMarketplace = selectedIDEs.some(id => id !== 'claude-code')`
scoped *only* to `copyPluginToMarketplace`, `runNpmInstallInMarketplace`,
and the pre-install `shutdownWorkerAndWait` (also pointless for claude-code-
only flows since we're not overwriting the worker's running cache dir
source). All other tasks (cache copy, register, enable, runtime setup) stay
unconditional.
**Claude Code missing → silent install of an IDE that isn't there**
When the user picked claude-code on a machine without it (e.g. a fresh
container), the install completed but `claude` was unavailable and the only
hint was a generic warn line. Replace with an explicit pre-flight prompt:
Claude Code is not installed. Claude-mem works best in Claude Code, but
also works with the IDEs below.
? Install Claude Code now?
◆ Yes — install Claude Code (recommended)
◯ No — pick another IDE below
◯ Cancel installation
If the user picks "Yes", run `curl -fsSL https://claude.ai/install.sh | bash`
(or the PowerShell equivalent on Windows), then re-detect IDEs and proceed
with claude-code pre-selected. If the install fails or the user picks "No",
the multiselect still appears with claude-code visible (just unmarked
[detected]), so they can opt in or pick another IDE.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(install): detect Claude Code via `claude` CLI, not ~/.claude dir
The directory `~/.claude` can exist (e.g. mounted in Docker, or created
by tooling) without Claude Code actually being installed. Detect the
`claude` command in PATH instead so the installer correctly offers to
install Claude Code when missing.
* docs(learn-codebase): add reviewer note explaining the cost tradeoff
The skill intentionally reads every file in full to build a cognitive
cache that pays off across the rest of the project. Add a brief note
so reviewers (human or bot) understand the tradeoff before flagging
the unbounded read as a cost issue.
* fix: address Greptile P1 feedback on welcome hint and learn-codebase
- SearchRoutes: skip welcome hint when caller passes ?full=true so
explicit full-context requests aren't intercepted by the hint.
- learn-codebase: replace `sed` instruction with the Read tool's
offset/limit parameters, since Bash is gated in Claude Code by
default.
* feat(install): ASCII-animated logo splash on interactive install
Plays a ~1s bloom animation of the claude-mem sunburst logomark when
the installer starts in an interactive terminal — geometrically rendered
via 12 ray curves around a center disc, in the brand orange. The
wordmark and tagline type on alongside the final frame.
Auto-skipped on non-TTY, in CI, when NO_COLOR or CLAUDE_MEM_NO_BANNER
is set, or when the terminal is too narrow.
Inspired by ghostty +boo.
* feat(banner): replace rotation frames with angular-sector bloom generator
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(banner): replace rotation frames with angular-sector bloom generator
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(banner): three-act choreography renderer with radial gradient and diff redraw
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(banner): update preview script to support small/medium/hero tier selection
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(docker): add COLORTERM=truecolor to test-installer sandbox
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(install): auto-apply PATH for Claude Code with spinner UX
The Claude Code install.sh prints a Setup notes block telling users to
manually edit "your shell config file" to add ~/.local/bin to PATH —
which left fresh installs unable to launch claude from the command line.
After a successful install, detect ~/.local/bin/claude on disk and, if
the dir is missing from PATH, append the right export line to .zshrc /
.bash_profile / .bashrc / fish config (idempotent, marked with a
comment). Also updates process.env.PATH for the current install run.
Wraps the curl|bash install in a clack spinner (interactive only) so the
~4 minute native-build download doesn't look frozen — output is captured
silently and dumped on failure for debuggability. Non-interactive mode
keeps inherited stdio for CI logs.
Verified end-to-end in the test-installer docker sandbox: spinner
animates, .bashrc gets the export, fresh login shell resolves claude.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(banner): video-frame ASCII renderer with three-act choreography
Generator switched from a single Jimp-rendered logo to pre-extracted
video frames concatenated with \x01 separators and gzip-deflated, ported
from ghostty's boo wire format. Renderer rewritten around three acts
(ignite → stagger bloom → text reveal + breathe) with adaptive sizing,
radial gradient, and diff-based redraw.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(onboarding): unify install / SessionStart / viewer around one first-success moment
Three surfaces now point at the same north-star moment — open the viewer, do
anything in Claude Code, watch an observation appear within seconds — with the
same verbatim timing and privacy lines, and a single canonical "how it works"
explainer instead of three diverging copies.
- Canonical explainer at src/services/worker/onboarding-explainer.md served via
GET /api/onboarding/explainer; mirrored into plugin/skills/how-it-works/SKILL.md
- SessionStart welcome hint rewritten as third-person status (no imperatives
Claude tries to execute), pinned with a default-value regression test
- Post-install Next Steps reframed as "two paths": passive default + optional
/learn-codebase front-load; drops /mem-search and /knowledge-agent from this
surface; adds verbatim timing + privacy lines and /how-it-works link
- /api/stats response gains firstObservationAt for the viewer stat row
- Viewer WelcomeCard branches on observationCount === 0: empty state shows live
worker-connection dot + "waiting for activity"; has-data state shows
observations · projects · since [date] and two example prompts. v2 dismiss key
- jimp added to package.json to fix pre-existing banner-frame build break
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(banner): play unconditionally; only honor CLAUDE_MEM_NO_BANNER
The 128-col / TTY / CI / NO_COLOR gates silently swallowed the banner in
narrower terminals, CI logs, and any non-TTY pipe — including Docker runs
where -it should preserve the experience but column width was the wrong
gate. Remove the implicit gates; keep the explicit opt-out only.
If a frame wraps in a narrow terminal, that's better than the banner
not playing at all.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* revert(banner): restore 15:33 gating logic per user request
Reverts eb6fc157. Restores isBannerEnabled to the state at commit
8e448015 (2026-04-30 15:33): TTY check, !CI, !NO_COLOR, !CLAUDE_MEM_NO_BANNER,
and cols >= BANNER.width.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(install): wrap remaining slow steps with spinners
Each IDE installer (Cursor, Gemini CLI, OpenCode, Windsurf, OpenClaw,
Codex CLI, MCP integrations) now runs inside a clack task spinner with
per-step progress messages instead of silent dynamic-import + cpSync.
Pre-overwrite worker shutdown (up to 10s) and the post-install health
probe (up to 3s) also get spinners.
Internal console.log/error/warn from each IDE installer is buffered
during the spinner; if the install fails, captured output is replayed
afterward via log.warn so users can see what broke.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(review): observation count + IDE pre-selection regressions
WelcomeCard's "no observations yet" empty state was triggered when a
project filter narrowed the feed to zero rows, even with thousands of
observations elsewhere. Source the count from global stats.database
to match firstObservationAt's scope.
Restore initialValues: [] in the IDE multiselect — pre-selecting every
detected IDE was the exact regression #2106 was filed for.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(install): trichotomy worker state + cache fallback for script path
ensureWorkerStarted now returns 'ready' | 'warming' | 'dead' instead of
boolean. The spawned-but-still-warming case (common in Docker cold
starts and slow first-time inits) was being misreported as 'did not
start', which contradicted the next-steps panel saying 'still starting
up'. Install task message and Next Steps headline now agree on the
actual state.
Also fixes the actual root cause of 'Worker did not start' on
claude-code-only installs: the worker script path was hardcoded to the
marketplace dir, which is left empty when no non-claude-code IDE is
selected. Now falls back to pluginCacheDirectory(version) when the
marketplace copy isn't present.
Verified end-to-end in docker/claude-mem with --ide claude-code,
--ide cursor, and a fresh container — install task and headline
agree on 'Worker ready at http://localhost:<port>' in all cases.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: align CLAUDE.md and public docs with current code
Sweep across CLAUDE.md and 10 high-traffic docs/public/ MDX files to
remove point-in-time references and align with the actual current
shape of the codebase. Highlights:
- Hardcoded port 37777 → per-user formula (37700 + uid % 100) on the
front-door pages (introduction, installation, configuration,
architecture/overview, architecture/worker-service, troubleshooting,
hooks-architecture, platform-integration).
- Default model 'sonnet' → 'claude-haiku-4-5-20251001' (matches
SettingsDefaultsManager).
- Node 18 → 20 (matches package.json engines).
- Lifecycle hook count corrected (5 events).
- Removed the nonexistent 'Smart Install' component and pre-built
directory tree referencing files that no longer exist
(context-hook.ts, save-hook.ts, cleanup-hook.ts, etc.); replaced
with the real worker dispatcher shape.
- Removed CLAUDE.md '#2101' issue tag (kept the design rationale).
- Replaced obsolete hooks.json example with a description of the real
bun-runner.js / worker-service.cjs hook event shape.
Lower-traffic doc pages still hardcode 37777 — left for a separate
global pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(scripts): land strip-comments around real parsers (postcss, remark, parse5)
Each language gets a real parser to locate comments, then we splice ranges
out of the original source. The library never serializes — that's how
remark-stringify produced 243 reformat-noise diffs in the first attempt
versus the 21 real strip targets here.
JS/TS/JSX -> ts.createSourceFile + getLeadingCommentRanges
CSS/SCSS -> postcss.parse + walkComments + node.source offsets
MD/MDX -> remark-parse (+ remark-mdx) + AST html / mdx-expression nodes
HTML -> parse5 with sourceCodeLocationInfo
shell/py -> kept hand-rolled hash stripper (no library worth the dep)
Preserves: shebangs, @ts-* directives, eslint-disable, biome-ignore,
prettier-ignore, triple-slash refs, webpack magic, /*! license keep,
@strip-comments-keep file marker. JS/TS handler runs a parse-roundtrip
check and refuses to write if syntax errors increased (catches the
worker-utils.ts breakage class from the 2026-04-29 attempt).
npm scripts:
strip-comments (apply)
strip-comments:check (CI-style, exits non-zero if changes needed)
strip-comments:dry-run (list, no writes)
Verified --check on this repo: 21 changes, -4.0% bytes, no parse-error
regressions, no reformat-suspect false positives.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor: strip comments codebase-wide via parser-backed tool
21 files changed, -17,550 bytes (-4.0%) of narrative comments removed
across .ts / .tsx / .js / .mjs and the .gitignore. JS/TS comments stripped
via ts.createSourceFile + getLeadingCommentRanges — same canonical lexer,
same behavior as the 2026-04-29 strip, no reformat noise.
Preexisting baseline (unchanged):
typecheck: 16 errors at HEAD, 16 errors after strip (line numbers shift,
no new error classes — verified via diff of sorted error lists)
build: fails at HEAD with CrushHooksInstaller.js unresolved import
(preexisting, unrelated to this strip)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(install): drop Crush integration references after extract
The Crush integration was extracted to its own branch on May 1, but the
import at install.ts:280 (and the case block + ide-detection entry +
McpIntegrations config + npx-cli help text) still referenced the now-
removed CrushHooksInstaller.js, breaking the build.
Removes:
- case 'crush' block in install.ts
- crush entry in ide-detection.ts
- CRUSH_CONFIG and registration in McpIntegrations.ts
- 'crush' from the IDE Identifiers help line in index.ts
Rebuilds worker-service.cjs to match.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(banner): mark generated banner-frames.ts with @strip-comments-keep
Without this, every build/strip cycle ping-pongs five lines of doc
comments in and out of the auto-generated output. The keep-marker tells
strip-comments.ts to skip the file entirely.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(build): drop banner-frame regen from build script
generate-banner-frames.mjs requires PNG frames in /tmp/cmem-banner-frames
that only exist after the maintainer runs ffmpeg locally on the source
video. CI has neither the video nor the frames, so the build broke on
Windows. The output (src/npx-cli/banner-frames.ts) is committed, so the
regen is a one-shot dev step — not a build step. Run the script directly
when the video changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(worker): unstick the spinner — kill claim-self-lock, wake on fail, auto-broadcast
Three surgical changes that cure the stuck-spinner bug at the source.
Phase 1.1 (L9): claimNextMessage no longer self-excludes its own worker_pid.
A single UPDATE-RETURNING grabs the oldest pending row by id. Removes the
LiveWorkerPidsProvider plumbing that was never injected — Supervisor enforces
single-worker via PID file, so the multi-worker SQL was defending against a
configuration the project does not support.
Phase 1.2 (L19): SessionManager.markMessageFailed wraps PendingMessageStore.markFailed
and emits 'message' on the per-session EventEmitter. The iterator's waitForMessage
now wakes immediately on re-pend instead of parking for 3 minutes. ResponseProcessor
and SessionRoutes routed through the new wrapper.
Phase 1.3 (L24): PendingMessageStore takes an optional onMutate callback fired
from every mutator (enqueue, claimNextMessage, confirmProcessed, markFailed,
transitionMessagesTo, clearFailedOlderThan). SessionManager wires it; WorkerService
passes broadcastProcessingStatus. Ten manual broadcast calls deleted across
SessionCleanupHelper, SessionEventBroadcaster, SessionRoutes, DataRoutes, and
worker-service. Caller discipline becomes structurally impossible to forget.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(worker): delete dead code — legacy routes, processPendingQueues, decorative guards
Pure deletions. Phase 2 of kill-the-asshole-gates.
- Legacy /sessions/:sessionDbId/* routes (handleSessionInit, handleObservations,
handleSummarize, handleSessionStatus, handleSessionDelete, handleSessionComplete)
bypassed all five ingest gates and were a parallel write path. Folded the
initializeSession + broadcastNewPrompt + syncUserPrompt + ensureGeneratorRunning
+ broadcastSessionStarted work into the canonical /api/sessions/init handler so
the hook makes one round trip instead of two.
- processPendingQueues (~104 lines, zero callers) — replaced in Phase 6 by a
one-statement startup sweep.
- spawnInProgress Map and crashRecoveryScheduled Set — decorative dedupe over
generatorPromise and stillExists checks that already provide the real safety.
- STALE_GENERATOR_THRESHOLD_MS — pre-empted live generators and raced with the
finally block; the 3min idle timeout already kills zombies.
- MAX_SESSION_WALL_CLOCK_MS — ran a SELECT on every observation to enforce 24h.
Runaway-spend protection lives in the API key, not in claude-mem.
- Missing-id 400 in shared.ts ingestObservation — Zod already enforces min(1)
on contentSessionId and toolName at the route schema.
- SessionCompletionHandler import + completionHandler field on SessionRoutes
(orphaned after handler deletions).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(worker): SQL-backed getTotalQueueDepth — single source of truth
Was: iterate this.sessions.values() and sum getPendingCount per session.
Now: SELECT COUNT(*) FROM pending_messages WHERE status IN ('pending','processing').
The in-memory sessions Map drifted from the DB rows whenever a generator exited
without confirm/fail, leading to false-positive isProcessing in the UI. Phase 1.3's
auto-broadcast fires on every mutation, but it broadcast a stale Map count.
Reading from the DB makes the UI's spinner state match what the queue actually holds.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(worker): typed abortReason replaces wasAborted boolean
Was: a boolean wasAborted that lumped every abort together. The finally block
branched on !wasAborted, so any abort skipped restart — including idle aborts
with pending work, which is exactly the case where we DO want to restart.
Now: ActiveSession.abortReason is a typed enum 'idle' | 'shutdown' | 'overflow'
| 'restart-guard'. The finally block consumes the reason and only skips restart
for 'shutdown' and 'restart-guard'. Idle and overflow aborts fall through, so
if pending work exists they trigger restart correctly.
Dropped 'stale' and 'wall-clock' from the union — Phase 2 deleted those paths.
Natural-completion abort (post-success) intentionally has no reason; it's not
gating restart logic.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(worker): unify the two generator-exit finally blocks
Was: worker-service.ts:startSessionProcessor and SessionRoutes:ensureGeneratorRunning
each had their own ~70-line finally block with divergent restart-guard handling.
The worker-service path called terminateSession on RestartGuard trip and orphaned
pending rows (the L16 bug); the SessionRoutes path drained them. Two places to
update when rules changed.
Now: handleGeneratorExit in src/services/worker/session/GeneratorExitHandler.ts
owns the contract:
1. Always kill the SDK subprocess if alive.
2. Always drain processingMessageIds via sessionManager.markMessageFailed
(which wakes the iterator — Phase 1.2).
3. shutdown / restart-guard reasons: drain pending rows via
transitionMessagesTo('failed'), finalize, remove from Map. Fixes L16.
4. pendingCount=0: finalize normally and remove from Map.
5. pendingCount>0: backoff respawn via per-session respawnTimer (no global Set;
Phase 2.4 deleted that). RestartGuard trip drains to 'abandoned'.
Both finally blocks are now ~10-line wrappers that translate local state into the
canonical abortReason and delegate. Restored completionHandler injection into
SessionRoutes (was dropped in Phase 2 cleanup; needed by the unified helper for
finalizeSession).
Behavior change: SessionRoutes' previous "keep idle session in memory" was
deliberately replaced by the plan's "remove from Map on natural completion" —
next observation reinitializes via getMessageIterator → initializeSession.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(worker): startup orphan sweep — reset 'processing' rows at boot
When the worker dies (crash, kill, restart), any pending_messages rows it left
in 'processing' state are by definition orphans (the only worker is dead).
Single SQL UPDATE at boot resets them to 'pending' so the iterator can claim
them again. Replaces the deleted processPendingQueues function (Phase 2.2).
Runs in initializeBackground after dbManager.initialize() and before the
initializationComplete middleware releases blocked HTTP requests, so no
in-flight request can race the sweep. NOT on a periodic timer — after boot,
every 'processing' row has a live consumer and a periodic sweep would race.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(worker): simplify enqueue catch, replace memorySessionId throw with re-pend
7.1: queueObservation's catch was logging two ERROR-level messages and rethrowing.
The rethrow is correct (FK violations / disk full / schema drift should crash
loudly), but the verbose ERROR logging pretended the error was recoverable.
Reduced to one INFO line + rethrow.
7.2: ResponseProcessor's memorySessionId guard was throwing if the SDK hadn't
included session_id on the first user-yield, terminal-failing the entire batch.
Now warns and re-pends in-flight messages via sessionManager.markMessageFailed
(which wakes the iterator — Phase 1.2). The next iteration tries again with
memorySessionId hopefully captured.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(sync): mirror builds to installed-version cache for hot reload
When package.json bumps past Claude Code's installed pin, sync-marketplace
wrote new code to cache/<buildVersion>/ but the worker loaded from
cache/<installedVersion>/, so worker:restart reloaded the same old code.
Replace the exit-on-mismatch preflight with a mirror step: when versions
differ, also rsync plugin/ into cache/<installedVersion>/ so worker:restart
hot-reloads new code without a Claude Code session restart. The
build-version cache still gets written for the eventual
`claude plugin update`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: delete dead barrel files and orphan utilities
- src/sdk/index.ts (re-exports parser+prompts; nothing imported the barrel)
- src/services/Context.ts (re-exports ./context/index.js; no importers)
- src/services/integrations/index.ts (no importers)
- src/services/worker/Search.ts (3-line barrel of ./search/index.js)
- src/services/infrastructure/index.ts: drop CleanupV12_4_3 re-export
- src/utils/error-messages.ts (getWorkerRestartInstructions never imported)
- src/types/transcript.ts (170 LoC of types, zero importers)
- src/npx-cli/_preview.ts (banner dev preview, no script wires it)
Build + tests still pass; observations still flowing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(parser): drop unused detectLanguage
Only the user-grammar-aware variant detectLanguageWithUserGrammars()
is actually called.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(types): drop unused SdkSessionRecord + ObservationWithContext
Both interfaces in src/types/database.ts had zero importers anywhere
in src or tests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(npx-cli): drop unused getDetectedIDEs + claudeMemDataDirectory
getDetectedIDEs has no callers — install.ts uses detectInstalledIDEs
directly. claudeMemDataDirectory has no callers either.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(ProcessManager): drop dead orphan-reaper + signal-handler helpers
Each had zero callers in src/ or tests/:
- cleanupOrphanedProcesses + enumerateOrphanedProcesses
- ORPHAN_PROCESS_PATTERNS + ORPHAN_MAX_AGE_MINUTES
- forceKillProcess
- waitForProcessesExit
- createSignalHandler
- resetWorkerRuntimePathCache
The orphan reaper was retired in PATHFINDER Plan 02 ("OS process groups
replace hand-rolled reapers", commit
|
||
|
Alex Newman
|
97c7c999b1 |
feat: basic claude-mem Docker container for easy spin-up (#2076)
* feat(evals): SWE-bench Docker scaffolding for claude-mem resolve-rate measurement Adds evals/swebench/ scaffolding per .claude/plans/swebench-claude-mem-docker.md. Agent image builds Claude Code 2.1.114 + locally-built claude-mem plugin; run-instance.sh executes the two-turn ingest/fix protocol per instance; run-batch.py orchestrates parallel Docker runs with per-instance isolation; eval.sh wraps the upstream SWE-bench harness; summarize.py aggregates reports. Orchestrator owns JSONL writes under a lock to avoid racy concurrent appends; agent writes its authoritative diff to CLAUDE_MEM_OUTPUT_DIR (/scratch in container mode) and the orchestrator reads it back. Scaffolding only — no Docker build or smoke test run yet. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(evals): OAuth credential mounting for Claude Max/Pro subscriptions Skips per-call API billing by extracting OAuth creds from host Keychain (macOS) or ~/.claude/.credentials.json (Linux) and bind-mounting them read-only into each agent container. Creds are copied into HOME=$SCRATCH/.claude at container start so the per-instance isolation model still holds. Adds run-batch.py --auth {oauth,api-key,auto} (auto prefers OAuth, falls back to API key). run-instance.sh accepts either ANTHROPIC_API_KEY or CLAUDE_MEM_CREDENTIALS_FILE. smoke-test.sh runs one instance end-to-end using OAuth for quick verification before batch runs. Caveat surfaced in docstrings: Max/Pro has per-window usage limits and is framed for individual developer use — batch evaluation may exhaust the quota or raise compliance questions. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(docker): basic claude-mem container for ad-hoc testing Adds docker/claude-mem/ with a fresh spin-up image: - Dockerfile: FROM node:20 (reproduces anthropics/claude-code .devcontainer pattern — Anthropic ships the Dockerfile, not a pullable image); layers Bun + uv + locally-built plugin/; runs as non-root node user - entrypoint.sh: seeds OAuth creds from CLAUDE_MEM_CREDENTIALS_FILE into $HOME/.claude/.credentials.json, then exec's the command (default: bash) - build.sh: npm run build + docker build - run.sh: interactive launcher; auto-extracts OAuth from macOS Keychain (security find-generic-password) or ~/.claude/.credentials.json on Linux, mounts host .docker-claude-mem-data/ at /home/node/.claude-mem so the observations DB survives container exit Validated end-to-end: PostToolUse hook fires, queue enqueues, worker's SDK compression runs under subscription OAuth, observations row lands with populated facts/concepts/files_read, Chroma sync triggers. Also updates .gitignore/.dockerignore for the new runtime-output paths. Built plugin artifacts refreshed by the build step. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(evals/swebench): non-root user, OAuth mount, Lite dataset default - Dockerfile.agent: switch to non-root \`node\` user (uid 1000); Claude Code refuses --permission-mode bypassPermissions when euid==0, which made every agent run exit 1 before producing a diff. Also move Bun + uv installs to system paths so the non-root user can exec them. - run-batch.py: add extract_oauth_credentials() that pulls from macOS Keychain / Linux ~/.claude/.credentials.json into a temp file and bind- mounts it at /auth/.credentials.json:ro with CLAUDE_MEM_CREDENTIALS_FILE. New --auth {oauth,api-key,auto} flag. New --dataset flag so the batch can target SWE-bench_Lite without editing the script. - smoke-test.sh: default DATASET to princeton-nlp/SWE-bench_Lite (Lite contains sympy__sympy-24152, Verified does not); accept DATASET env override. Caveat surfaced during testing: Max/Pro subscriptions have per-window usage limits; running 5 instances in parallel with the "read every source file" ingest prompt exhausted the 5h window within ~25 minutes (3/5 hit HTTP 429). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: address PR #2076 review comments - docker/claude-mem/run.sh: chmod 600 (not 644) on extracted OAuth creds to match what `claude login` writes; avoids exposing tokens to other host users. Verified readable inside the container under Docker Desktop's UID translation. - docker/claude-mem/Dockerfile: pin Bun + uv via --build-arg BUN_VERSION / UV_VERSION (defaults: 1.3.12, 0.11.7). Bun via `bash -s "bun-v<V>"`; uv via versioned installer URL `https://astral.sh/uv/<V>/install.sh`. - evals/swebench/smoke-test.sh: pipe JSON through stdin to `python3 -c` so paths with spaces/special chars can't break shell interpolation. - evals/swebench/run-batch.py: add --overwrite flag; abort by default when predictions.jsonl for the run-id already exists, preventing accidental silent discard of partial results. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: address coderabbit review on PR #2076 Actionable (4): - Dockerfile uv install: wrap `chmod ... || true` in braces so the trailing `|| true` no longer masks failures from `curl|sh` via bash operator precedence (&& binds tighter than ||). Applied to both docker/claude-mem/ and evals/swebench/Dockerfile.agent. Added `set -eux` to the RUN lines. - docker/claude-mem/Dockerfile: drop unused `sudo` apt package (~2 MB). - run-batch.py: name each agent container (`swebench-agent-<id>-<pid>-<tid>`) and force-remove via `docker rm -f <name>` in the TimeoutExpired handler so timed-out runs don't leave orphan containers. Nitpicks (2): - smoke-test.sh: collapse 3 python3 invocations into 1 — parse the instance JSON once, print `repo base_commit`, and write problem.txt in the same call. - run-instance.sh: shallow clone via `--depth 1 --no-single-branch` + `fetch --depth 1 origin $BASE_COMMIT`. Falls back to a full clone if the server rejects the by-commit fetch. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: address second coderabbit review on PR #2076 Actionable (3): - docker/claude-mem/run.sh: on macOS, fall back to ~/.claude/.credentials.json when the Keychain lookup misses (some setups still have file-only creds). Unified into a single creds_obtained gate so the error surface lists both sources tried. - docker/claude-mem/run.sh: drop `exec docker run` — `exec` replaces the shell so the EXIT trap (`rm -f "$CREDS_FILE"`) never fires and the extracted OAuth JSON leaks to disk until tmpfs cleanup. Run as a child instead so the trap runs on exit. - evals/swebench/smoke-test.sh: actually enforce the TIMEOUT env var. Pick `timeout` or `gtimeout` (coreutils on macOS), fall back to uncapped with a warning. Name the container so exit-124 from timeout can `docker rm -f` it deterministically. Nitpick from the same review (consolidated python3 calls in smoke-test.sh) was already addressed in the prior commit ef621e00. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: address third coderabbit review on PR #2076 Actionable (1): - evals/swebench/smoke-test.sh: the consolidated python heredoc had competing stdin redirections — `<<'PY'` (script body) AND `< "$INSTANCE_JSON"` (data). The heredoc won, so `json.load(sys.stdin)` saw an empty stream and the parse would have failed at runtime. Pass INSTANCE_JSON as argv[2] and `open()` it inside the script instead; the heredoc is now only the script body, which is what `python3 -` needs. Nitpicks (2): - evals/swebench/smoke-test.sh: macOS Keychain lookup now falls through to ~/.claude/.credentials.json on miss (matches docker/claude-mem/run.sh). - evals/swebench/run-batch.py: extract_oauth_credentials() no longer early-returns on Darwin keychain miss; falls through to the on-disk creds file so macOS setups with file-only credentials work in batch mode too. Functional spot-check of the parse fix confirmed: REPO/BASE_COMMIT populated and problem.txt written from a synthetic INSTANCE_JSON. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |