mirror of https://github.com/garrytan/gstack.git
3 Commits
| Author | SHA1 | Message | Date |
|---|---|---|---|
Garry Tan
|
1d41ee3ab3
|
v1.63.0.0 feat: GStack 2 fork port wave — egress receipts, context-bill, sharded gate, /health fix (#2541)
* test(helpers): shared skill-census helper with three explicit counts
physicalSkillFiles (symlinked dirs included, root router included),
authoredSkills (realpath-deduped, router excluded), registryEntries
(what ./setup registers: unique frontmatter names + _gstack-command).
One counting authority for the hermetic seeder, context-bill ground
truth, and the catalog-budget test — connect-chrome's dir symlink and
the root router otherwise produce three subtly different hand-rolled
censuses. Ported-wave foundation (C11).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(evals): stop the harness grading itself
findPreviousRun excluded only the file being written, by name, so every
suite compared against _partial-e2e.json — the current run's own
accumulator, relabelled with the current tier just before each flush.
That is why every block read '+$0.00, +0s, Stable run, no regressions.'
This harness has never been able to detect a regression, and reassuring
output that cannot fail is worse than none. In-progress runs are now
excluded by role, and a run with nothing to compare against says NO
BASELINE instead of claiming stability.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit f3140b5245221fff7fb9411c7ec07c2ca11587b5)
* refactor(evals): shared partial-run predicate + finalized-run lookup
isPartialEval(data, filename) is the one place that decides what counts
as an in-progress accumulator (the _partial flag OR a _partial-prefixed
filename), and findLatestFinalizedRun(evalDir, tier) is the one place
that finds the newest real run — scanning the eval dir plus one level of
shards/<slug>/ subdirs, where the sharded paid runner points each
shard's collector. skill-budget-regression.test.ts's hand-rolled
findLatestRun (flag-blind: a flagged-but-renamed accumulator passed its
name check) is replaced by the shared helper.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit b55fcf6966366fd21a8cdc46de61aab6e1b1d100)
* feat(evals): register shipped skills for hermetic PTY children
Hermetic children get a config dir that deliberately seeds no skills —
right for children that install their own, fatal for the PTY family that
TYPES /office-hours or /plan-ceo-review: claude rejects the command as
Unknown before any model turn, so the plan-family gate smokes measure
nothing. hermeticSkillsConfigDir() is a second, opt-in config dir under
the same runRoot that mirrors ./setup's registration exactly (real dir
per registry name, SKILL.md + sections/ symlinks, frontmatter-name
resolution, _gstack-command root alias), driven by the shared
skill-census so connect-chrome's dir symlink collapses the same way
setup's idempotent overwrite does.
Ported from fork commit 03c4eca2, tree walk rewritten for the upstream
layout (top-level <skill>/SKILL.md dirs, no skills/ tree). Unit tests
are new: seed shape, census parity, symlink resolution, connect-chrome
collapse, idempotence, no-API-key seed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 93dae6107b30ce453a07c2d342b60262bba6ce0b)
* feat(evals): seedSkills opt-in for PTY slash-command tests + tripwire
Wire ClaudePtyOptions.seedSkills through launchClaudePty: when set (and
hermetic, and no per-test CLAUDE_CONFIG_DIR override), the child gets
hermeticSkillsConfigDir() so typed /skill slash commands resolve instead
of dying as Unknown command before any model turn. Opted in at the three
runPlanSkill* helpers and the four direct-launch slash-command tests
(plan-design-with-ui, plan-ceo-mode-routing, autoplan-chain,
ship-idempotency).
New static tripwire (test/pty-skill-seeding-wiring.test.ts): any test
file that sends a slash command over the PTY must route through a
runPlanSkill* helper or pass seedSkills: true — an unseeded slash-command
test spends money and measures nothing. hermetic-wiring.test.ts now
blesses the repo-tree seeding path explicitly (config dir under runRoot,
symlinks into the repo checkout, never operator ~/.claude).
The CI "Register gstack skills for PTY smoke" step keeps a keep-me note:
container cross-mount symlinks defeat the TUI scanner and HOME is not
hermeticized, so the real-file copies there must survive this change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 63c52269daaffb833b3105ea9b4b99be6df8fec7)
* refactor(evals): single shared paid-test-set module
test/helpers/paid-test-set.ts is now the one definition of which test
files are paid (the exact globs package.json's test:gate expands).
scripts/test-free-shards.ts derives its free/paid exclusion from it
instead of a private regex list, dropping the dead
browse/test/security-review-fullstack.test.ts pattern (file no longer
exists). The sharded paid runner derives its enumeration from the same
module, so a file added to one list can no longer silently miss the
other.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit a7f36479a6a1f3656452370f5883371f3cb65623)
* feat(evals): env-driven lazy eval dir + shard-aware store and tooling
Importing eval-store no longer spawns the gstack-slug subprocess: the
module-level DEFAULT_EVAL_DIR constant is now a memoized defaultEvalDir()
resolved at collector construction. Resolution order: explicit
constructor arg, then GSTACK_EVAL_DIR, then slug detection — so the
sharded paid runner can point each shard child at its own
<evalDir>/shards/<slug>/ dir with plain env, no --preload.
Runs collected under a shards/ subdir record their slug in the eval
JSON (EvalResult.shard). findPreviousRun scans one shards/<slug>/ level
and prefers same-slug priors, so each shard baselines against its own
history instead of whichever shard flushed last. eval:list,
eval:summary, and eval:compare enumerate the same one level of shard
subdirs; eval:compare's no-arg mode also stops picking an in-progress
accumulator as the after-run.
eval-watch stays flat (documented follow-up): it tails a single dir for
live progress and gains nothing from per-shard baselines until the
runner emits a merged stream.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit e1f53f7d9c7fe6b65877d843f2e25bd2e2d12ffd)
* feat(evals): sharded paid tier runner
scripts/test-paid-shards.ts runs the gate/periodic tier one Bun process
per test file, with an EXTERNAL wall-clock timeout that SIGKILLs the
shard's detached process group and an aggregate that distinguishes
passed / failed / timed-out / never-started — partial execution can no
longer read as a pass. Bun's native --shard/--isolate covers none of
this: no process-group kill (hung claude/codex PTY grandchildren
survive in-process isolation), no never-started taxonomy, no per-shard
env. Each shard child gets GSTACK_EVAL_DIR=<evalDir>/shards/<slug>/
(slug = test filename sans extension, stable across runs) so shard
baselines compare against their own prior runs.
Output classification lives in scripts/test-strict-output.ts (strict
exit-code derivation, incremental fail-line classifier, child signal
forwarding) so the runner and any future strict bun-test wrapper share
one implementation. Enumeration derives from the shared paid-test-set
module; tier exclusion fires only on an explicit whole-file
EVALS_TIER === '<other>' guard.
package.json gains test:gate:sharded / test:periodic:sharded, and
eval:bg:gate / eval:bg:periodic now run the sharded scripts with detach
timeouts sized to the worst case (gate: 49 shards x 30min / 4 jobs ~
6.2h -> 25200s; periodic: 59 -> 28800s).
test/paid-shards.test.ts pins enumeration, tier classification, and the
kill-and-continue property with a real busy-loop shard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 5e76bd5931836257f896cedfe4e93912cb759c70)
* feat(security): hash-chained egress receipt ledger (core)
Port lib/egress-receipt from the v2 fork as TypeScript: writeReceipt
(sync, fail-closed via typed EGRESS_RECEIPT_FAILED), best-effort
writeOutcome, readLedger/listReceipts/verifyLedger, GSTACK_HOME ->
GSTACK_STATE_DIR -> ~/.gstack resolution, 0600 ledger under a 0700
security dir, and an mkdir spin lock (2.5s budget) with documented
>10s-mtime stale-lock reclaim.
Changes vs the fork:
- lastRawLine tail-reads the final 4KB instead of loading the whole
ledger, so appends stay O(1) as the file grows.
- WARN-at-size: past 25MB writeReceipt emits one self-explanatory
stderr warning per process (what the ledger is, how to inspect it,
rotation TODO); verifyLedger gains a sizeWarning field. Rotation
TODO carries the chain-genesis sketch (new generation's first record
embeds the prior file's tail hash).
bin/gstack-egress-receipt is a bun script bridging shell callers:
write|outcome subcommands, exit 3 + EGRESS_RECEIPT_FAILED on stderr on
failure; --no-payload records sha256:null for git-class ops.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 619726a3d77d987a2e50151a5727b3faaaf5fc6a)
* chore(bin): delete dead brain-consumer/reader scripts
bin/gstack-brain-consumer and bin/gstack-brain-reader are byte-identical
dead scripts that POST the repo URL + a Bearer token to a /ingest-repo
endpoint gbrain removed (docs/gbrain-sync.md already documents the
removal in past tense). No live references remain; CHANGELOG mentions
are historical.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 254ddc69fc5a0270fcc973e36b6a81766d835d2d)
* feat(security): shared shell receipt helpers
bin/gstack-egress-lib.sh (sourced library, gstack-gbrain-lib.sh
precedent) provides _receipted_curl and _receipted_git: write the
egress receipt BEFORE the send via gstack-egress-receipt, hand curl the
SAME payload file via --data-binary @file so the receipt hash matches
the wire bytes exactly, then append a best-effort outcome. Per-call
fail policy: 'closed' refuses the send (return 3, problem/cause/fix
message on stderr) and 'open' warns and proceeds. Payload temp files
are consumed immediately per call — no EXIT traps, since callers like
gstack-telemetry-sync own their own EXIT trap and a sourced trap would
clobber it.
Tested end-to-end against a local Bun.serve listener: receipt sha256
equals the sha256 of the bytes the listener received, fail-closed
refusal never touches the network and carries the problem/cause/fix
stderr shape, fail-open warns and proceeds.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 6d067dce2d4c8815dec98be551763c85a3671357)
* feat(security): receipt core shell sinks
Wire the three core bash egress sinks through gstack-egress-lib.sh:
- gstack-telemetry-sync: the batch POST now writes the payload to a
temp file, receipts those exact bytes fail-closed, and hands curl the
SAME file. On refusal nothing is sent and the cursor does not
advance, so the batch stays buffered for the next run. The HTTP
status is recorded as the receipt outcome.
- gstack-update-check: fail-open receipts (warn + proceed) on the
Supabase ping POST, both VERSION curls (via a local
_receipted_version_fetch helper that skips non-network schemes), and
git ls-remote. The ping receipt is written inside the backgrounded
subshell, so it can never block the script's exit.
- gstack-brain-sync: fail-closed git-class receipts. The push receipt
is written BEFORE the commit consumes the queue, so a refused receipt
leaves the queue intact and the next run retries the whole drain
(pinned by a new queue-intact-on-refusal test, including the
problem/cause/fix refusal message shape). The retry-path fetch and
retry push carry their own fail-closed receipts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 3c60f699acceaf1c92a218874711e05fc17dca5d)
* feat(security): receipt TS module sinks + tunnel
writeReceipt (fail-closed, sha256:null — a subprocess or SDK owns the
wire bytes) before every TS-module network-bearing operation:
- bin/gstack-gbrain-sync.ts: before the gbrain code walk that ships
repo content to the user's gbrain DB (may be remote Postgres). A
refused receipt fails the stage with status refused-egress-receipt.
- bin/gstack-memory-ingest.ts: before the gbrain batch import of
transcript pages. A refused receipt returns a system_error verdict
without spawning the import.
- browse/src/server.ts: before both ngrok.forward call sites (start-up
BROWSE_TUNNEL=1 path and the /tunnel/start endpoint). A receipt
failure lands in the existing catch that tears the tunnel listener
back down and refuses the start.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 5677d618a48fcd0ae2b068bf868781d90f809cb5)
* feat(design): receipted fetch for OpenAI calls
design/src/receipted-fetch.ts wraps every api.openai.com call: a
content-free egress receipt (sink design-openai, sha256 of the JSON
body — hash only, never the body) is written BEFORE the send. Polarity
is FAIL-OPEN: user-facing generation must not die because an audit log
hiccuped, so a receipt failure warns on stderr and the call proceeds.
Streams pass through untouched (response bodies returned as-is;
non-string request bodies receipted as sha256:null rather than drained
to hash).
All ten call sites converted with per-command payload classes:
generate, variants (injected fetchFn passes through), iterate (both
threaded and fresh paths), evolve (image + screenshot analysis), check,
diff, design-to-code, memory.
Unit-tested with injected fetch: receipt-before-send ordering, stream
passthrough, and fail-open on an unwritable ledger.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit c0e5ff6639414ac2fd98e8ac3affb51401746b55)
* feat(security): receipt admin scripts + user git-ops (zero exceptions)
Wire the remaining shell egress through gstack-egress-lib.sh:
- gstack-gbrain-mcp-verify: both JSON-RPC probe POSTs (initialize +
tools/list) receipted fail-closed via payload files (hash == wire
bytes). A refused receipt lands in the NETWORK class — no send.
- gstack-security-dashboard / gstack-community-dashboard: the
community-pulse GETs receipted fail-open (read-only stats must not
break over an audit hiccup).
- gstack-gbrain-supabase-provision: api_call receipted fail-closed.
Each retry attempt hands the helper a fresh copy of the body file
(the helper consumes its payload). The receipt hashes the request
body only — the PAT never reaches the ledger or any log. Refusal
exits 8 without retrying.
- git-class sha256:null receipts, fail-open: gstack-artifacts-init
(ls-remote, initial push, fetch/pull recovery, retry push),
gstack-brain-restore (staging clone, existing-repo fetch),
gstack-session-update (self-update pull).
gstack-team-init needs no wiring: every git clone in it is inside an
echoed instruction string, not an executed command.
The lib now self-locates with shell builtins only (no dirname), so
sourcing works under the whitelist-PATH test harnesses.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit b8c5e2055b21ab72878b3e46f8047782ee65a11c)
* test(security): egress wiring tripwire + polarity contract
Static-grep tripwire pinning the egress-receipt wiring (threat model in
the header: the ledger is forensic observability of ATTEMPTED egress,
not an exfiltration control):
- Per-sink assertions: every wired TS module imports egress-receipt and
calls writeReceipt; every wired shell sink sources
gstack-egress-lib.sh with each network op under a receipt;
ngrok-proximity check for server.ts; every design api.openai.com call
routes through receiptedFetch.
- Absence assertions: the dead brain-consumer/reader scripts stay
deleted (lstat, so a dangling symlink also fails).
- Polarity table pinned as data (fail-closed: brain-sync,
memory-ingest, gbrain-sync, telemetry-sync, ngrok, mcp-verify,
supabase-provision; fail-open: design-openai, update-check,
dashboards, git-class user ops, context-bill --exact) plus per-file
polarity spot-checks.
- NEW-SINK SCANNER with zero KNOWN_UNWIRED: sweeps bin/, lib/,
scripts/, design/src, browse/src for curl, absolute-URL fetch(, and
git remote ops (never local rev-parse/get-url; heredoc bodies and
message strings excluded) and requires every hit to be receipted or
in a REASONED exemption list where each entry carries its why.
Preamble-generated skill prose documented out-of-scope in the header.
- Shebang tripwire: no bin/gstack-* file may carry a node shebang.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit ff69ceeafaf9c017d539b6ad77ff8f95b680b979)
* feat(cli): gstack-egress reader
bin/gstack-egress (bun) — the auditor's view of the receipts ledger:
- list: one row per receipt (what gstack ATTEMPTED to send), with
--since/--host/--sink filters and --json.
- verify: recompute the hash chain; exit 3 on tamper naming the first
broken line; prints the sizeWarning when the ledger passes 25MB.
- grants: what CAN leave, built on the upstream config keys only
(telemetry, artifacts_sync_mode, redact_repo_visibility,
redact_prepush_hook via gstack-config get) — each grant names its
file, key, and the exact revoke command.
CLI smoke tests spawn the real bin against a temp GSTACK_HOME,
including a broken-chain fixture asserting exit 3.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 9e24eca0f1069fea2ea69e7df4e9b256e93d59a3)
* feat(cli): context-bill — token bill-of-materials (stripped port)
lib/context-bill.ts, ported from the v2 fork and STRIPPED to the tiers
this repo's skills can exercise: ALWAYS-ON (per-skill frontmatter bytes
with dead-key and foreign-host-file flags), EAGER (SKILL.md + any
forced 'for every invocation' references), on-disk totals, --diff,
--budget, and --exact with the calibration table. The fork's
CONDITIONAL/TRANSITIVE/LAZY/FAST-PATH parsers understand only its
dispatcher layout and were dropped; the tier fields stay in the report
shape (empty/zero/null) so re-adding a parser is additive.
TOKEN_DIVISORS and their provenance docblock kept; --help notes
recalibration via --exact's calibration block.
Three upstream fixes over the fork:
(a) findSkillDirs treats the walk ROOT as a container — the repo root's
router SKILL.md is billed AND its children are walked (the fork
short-circuited and billed one skill); walkMd skips node_modules
and dot-directories.
(b) installed-tree layout: subdirs that are their own repo checkout
(a gstack/ clone inside ~/.claude/skills, detected by .git) are
skipped, and directory symlinks (connect-chrome) are followed with
a container-recursion cycle guard.
(c) ROUTER_KEYS widened to the upstream frontmatter contract {name,
description, version, allowed-tools, triggers, preamble-tier}.
--exact writes an egress receipt (sink context-bill-exact, host
api.anthropic.com) BEFORE any count_tokens POST; if the receipt cannot
be written the run degrades to the offline estimate with a warning —
nothing is sent unrecorded. bin/gstack-context-bill is the bun shim.
Tests: fixture-tree ledgers, the three fixes, --diff/--budget exit
codes, --exact with injected fetch (envelope subtraction, receipt
ordering, fail-open degradation), CLI smoke test, and ground truth
against THIS repo via test/helpers/skill-census.ts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 675c19876b87ec927b555f5f64c7f93130b3de90)
* test(catalog): aggregate discovery-surface budget with ratchet protocol
Every host loads every skill's frontmatter name + description at
discovery, every session. applyCatalogTrim in scripts/gen-skill-docs.ts
shapes each description and the 160KB per-file warn covers body size,
but nothing capped the aggregate frontmatter — the catalog could grow
one reasonable-looking description at a time. This test is that
enforcement layer.
Measures the catalog via test/helpers/skill-census.ts authoredSkills
(symlink-deduped, root router counted separately as the _gstack-command
alias line item): 53 skills + router = 4,420 bytes = 1,105
token-equivalents today, asserted <= 1,150 (~4% headroom). Per-skill
sub-cap of 260 bytes (largest today: design-consultation at 229), plus
a non-empty-description check.
Failure messages are self-service ratchets: they print the new total,
the delta, and the update protocol (bump the constant AND the
derivation comment in the same commit; trim instead of grow for
existing descriptions). Parser handles folded block scalars
(description: >-) for fork parity; import-free by design so it
survives generator refactors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit c106fb36f768181b80c257e5cff1cde4f435f9c0)
* fix(browse): extension token bootstrap moves to pinned-origin POST; /health carries no token
GET /health is now liveness/status only in every mode — both token
carve-outs (headed-mode disjunct AND chrome-extension:// Origin
disjunct) are removed. Token bootstrap is POST /extension-token on the
local listener: the Origin header must be exactly
chrome-extension://<GSTACK_EXTENSION_ID> and the Host header's hostname
must parse to 127.0.0.1 or localhost (parsed via new URL, never literal
equality — Host arrives as '127.0.0.1:34567'). Wrong origin/host → 403
with no detail. The tunnel surface 404s the endpoint (not in
TUNNEL_PATHS, verified by test).
The extension ID is pinned by a new "key" field (RSA public key) in
extension/manifest.json; browse/scripts/extension-id.ts reproduces the
ID derivation (first 16 bytes of SHA-256 of the DER public key, hex
mapped 0-9a-f → a-p). The private key is not committed anywhere —
unpacked/baked-in loads only need the public key.
Extension side: background.js bootstraps and refreshes the token via
POST /extension-token (403 → disconnected state); sidepanel.js direct
connect path does the same; sidepanel-terminal.js's dead /health token
fallback (read AUTH_TOKEN/authToken keys the server never sent,
hardcoded port) is replaced with the window.gstackAuthToken path.
MIGRATION NOTE: the manifest key pins the extension ID, so existing
installs' side-panel local state (saved port, snoozes) resets once —
explained in-product via a one-time notice (flag
gstack_id_migrated_v162). After upgrading the server, restart the
browser so the old service worker stops polling for a token GET /health
no longer serves.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit e9a0b6847a2d17fe6656a4686b4efd0c8380eb09)
* docs: correct stale compiled-binaries claim; file three egress/eval follow-ups
CLAUDE.md's compiled-binaries section claimed browse/dist binaries are
tracked by git and appear as modified in git status — false since
|
|
|
Garry Tan
|
9fd03fae9e
|
v1.58.4.0 fix: high-priority community bug wave + PTY plan-mode smoke gate (#2077)
* fix(gbrain): stop forcing GBRAIN_PREPARE on transaction-mode poolers (#1965) buildGbrainEnv auto-set GBRAIN_PREPARE=true whenever DATABASE_URL targeted port 6543, and the /sync-gbrain capability check exported it for the rest of the skill run. Both had the semantics inverted: gbrain auto-disables prepared statements on transaction-mode poolers because they break every write there ("prepared statement does not exist"); GBRAIN_PREPARE=true is gbrain's documented override for SESSION-mode poolers on 6543, not a requirement for transaction mode. The #1435 search symptom the auto-set worked around was fixed gbrain-side. Remove both force-sets. A caller-set GBRAIN_PREPARE (either value) still passes through untouched, preserving the session-mode-on-6543 escape hatch. isTransactionModePooler stays exported. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(gbrain): classify probe timeout as its own status; sync proceeds instead of skipping (#1964) The 5s engine probe misclassified healthy-but-slow engines (cold Supabase pooler connections measured at 6.9-10.7s) as broken-config, so /sync-gbrain silently skipped code+memory and told the user their config was malformed. - New "timeout" status: probe killed at the deadline with no recognized stderr pattern. Default deadline is now 15s, overridable via GSTACK_GBRAIN_PROBE_TIMEOUT_MS (tests set 300ms against a fake that sleeps 2s). - Sync stages PROCEED on timeout with a stderr warning naming the env knob; a genuinely-dead engine surfaces its real error at the first operation instead of a false config diagnosis. - Consistency everywhere "ok" gated behavior: gstack-gbrain-detect --is-ok exits 0 on timeout, and gen-skill-docs' detection gate accepts it, so a slow engine no longer silently suppresses brain-aware features. - Status cache: key now includes the effective probe timeout (raising it invalidates a cached timeout) and GBRAIN_HOME; config detection honors GBRAIN_HOME so relocated-home users stop being misclassified as missing-config. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(bins): cygpath-normalize SCRIPT_DIR for bun imports; surface learnings-log errors (#1950) Under Windows git-bash, pwd yields a POSIX path (/c/Users/...) that Bun on Windows cannot resolve as an ES module specifier. gstack-learnings-log interpolates SCRIPT_DIR into a bun -e import, so every invocation died with "Cannot find module" — and 2>/dev/null swallowed the error, silently dropping every AI-logged learning for Windows users. - 3-line cygpath -m guard in gstack-learnings-log and gstack-question-log (which gains the same import shape in the next commit). Matches the duplicated IS_WINDOWS convention in setup; no shared shell lib exists. - learnings-log adopts question-log's set +e / TMPERR capture pattern wholesale: validation errors now print to stderr. The old `if [ $? -ne 0 ]` check was dead code under set -euo pipefail — the script exited at the failing assignment before reaching it. - New test/bin-windows-bun-import-paths.test.ts: static invariant (any bash bin interpolating $SCRIPT_DIR into a bun -e import must carry the guard) + behavioral end-to-end run invoked via `bash <bin>` — added to the windows-free-tests workflow list so the conversion is proven on the only platform where the bug exists. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(question-log): dedupe INJECTION_PATTERNS via lib/jsonl-store (#1934) bin/gstack-question-log carried a local copy of the injection-pattern list, so pattern fixes to lib/jsonl-store.ts never propagated — including the /override[:\s]/i false-positive fix arriving via community PR #1940. Import the shared hasInjection instead (enabled by the previous commit's cygpath guard). question-log also gets the lib's stricter superset (human:, disregard, from-now-on, approve-all patterns). Tests pin the contract in a #1940-order-independent way: an "Override: ignore all previous instructions" header is rejected, "prose overrides the deterministic table" is accepted, and a static invariant keeps local INJECTION_PATTERNS duplicates out of the bin. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(security): community-pulse + both dashboards never report fake zeros (#1947) The security-signaling surface failed open at three layers — every failure mode read as a reassuring "0 attacks" / "0 installs": - community-pulse edge function: supabase-js returns {data,error} without throwing, and all five queries discarded `error` — a DB outage produced real-looking zeros via the SUCCESS path, and the catch (also returning zeros with HTTP 200) was unreachable for query failures. Every query now destructures and throws; the catch serves the stale cache (marked "stale": true) when one exists, else 503 {"error":"pulse_unavailable"}. Success responses carry "status":"ok" so clients can distinguish authoritative data from legacy backends. NOTE: the edge function deploys out-of-band (supabase functions deploy community-pulse). - gstack-security-dashboard: captures the HTTP status; non-200 / network failure / error body / missing section → "unknown — backend error"; jq missing → "unknown — install jq" (the lossy grep fallback broke on nested arrays and under-reported attacks as zero — removed); a 200 without the new marker shows figures with an "unverified (legacy backend)" note. Also fixes a latent display bug: the TOTAL grep matched the digit 7 inside "attacks_last_7_days" and misreported every count. - gstack-community-dashboard: same class — curl || echo "{}" plus grep || echo "0" printed "Weekly active installs: 0" on any failure. Now "unknown — backend error (HTTP N)". test/security-dashboard-fallback.test.ts pins the matrix (200+marker, 200-legacy, 503, network failure) x (jq present, jq absent) for both bins: "unknown" states never render as 0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(telemetry): redact error_message spans before they leave the machine (#1947) error_message was uploaded with only quote/newline escaping — stack traces and failed-API errors can embed credentials, private paths, and hostnames, and the sync path strips only _repo_slug/_branch. New lib/redact-engine.ts export redactFindingSpans(): replaces EVERY finding's span with <REDACTED-{id}> regardless of tier (applyRedactions is the interactive PII-only path and exits nonzero on credential findings, so it can't serve machine egress). Returns null when a span can't be located — callers drop the whole payload rather than risk a leak. gstack-telemetry-log pipes error_message through it at LOG time, so the local JSONL at rest is clean too; surrounding text survives for crash triage. FAIL CLOSED: bun missing, engine error, or non-JSON-string output all null the field. Tests pin: embedded ghp_ token → <REDACTED-github.pat> with context intact; redactor unavailable → null; raw bytes on disk never contain the token. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(redact): prepush guard fails closed on git failure; /ship owns hook install (#1946) Two gaps closed: 1. Fail closed. The git() helper returned "" on ANY non-zero exit or maxBuffer overflow (status null), addedLinesFor produced an empty string, and the push sailed through unscanned — fail-open on exactly the oversized-diff case where a large secret-bearing blob is most likely. The diff call now uses a strict variant that throws; main blocks with a clear message naming the GSTACK_REDACT_PREPUSH=skip escape valve. Probe calls (symbolic-ref, rev-parse, merge-base) keep the permissive helper — their failures are normal control flow. 2. Install path. The hook was installed by nothing ("opt-in, installed by nothing" was the issue's words). ./setup runs in the gstack checkout — the wrong repo for a per-project hook — so it gets a one-line hint only. /ship owns per-repo install: config redact_prepush_hook=true + hook missing → silent install (consent already given); config unset + no ~/.gstack/.redact-prepush-prompted marker → one-time machine-wide AskUserQuestion offer, answer persisted. ship/SKILL.md regenerated in this same commit (check-freshness bisect discipline). Tests: unscannable diff (bogus SHAs) → exit 1 + valve named; empty-but- successful diff → exit 0; static asserts pin setup as hint-only and the ship template as the installer surface. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(redact): six new credential patterns — GitLab, HuggingFace, npm, DigitalOcean, Bearer, GCP SA (#1946) Coverage gaps from the #1946 security review, including token types for tooling gstack itself drives (glab): HIGH (block): gitlab.token (glpat-/glptt-/gldt-), huggingface.token (hf_), npm.token (npm_), digitalocean.token (dop_v1_), gcp.service_account (the JSON-escaped "private_key" form that dodges pem.private_key's literal-block match when minified, confirmed by "private_key_id" proximity). MEDIUM (warn): auth.bearer — the most FP-prone shape in the set (docs are full of "Authorization: Bearer <token>"), so it requires header-context proximity and the same entropy>=3.0 + placeholder validator recipe as env.kv. "Bearer YOUR_TOKEN_HERE" never fires; calibration over coverage, per the cries-wolf principle. All shapes are linear-time; test/redact-pattern-lint.test.ts covers them automatically. Engine tests add positive + placeholder-negative cases per pattern. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test: coverage-audit additions for the fix wave Ship Step 7 gap-fill (all passing, 248 tests across the touched suites): memory + dream stage probe-timeout proceeds, gbrain-detect override paths, stale-flag passthrough, 200-body-missing-.security fail-closed case, telemetry redaction edges, and credential-pattern edge cases. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: pre-landing review fixes Review army findings (1 critical, auto-fixed with regression tests): - CRITICAL (security specialist, verified live): redactFindingSpans spliced only the regex capture span, and pem.private_key / gcp.service_account capture just the BEGIN-header — the key body survived "redaction" and shipped via telemetry. Marker-only patterns now drop the whole payload (null, fail closed). Overlapping spans (Bearer+JWT on the same bytes) are coalesced before splicing so stale offsets can't leave partial secret bytes behind. - gitStrict: drop the dead `|| r.status === null` disjunct (null !== 0 already covers it); add the signal-kill/null-status regression test the docstring promised. - security-dashboard human mode flags stale snapshots ("figures may be out of date") instead of presenting frozen counts as current. - community-dashboard marker check uses jq when available — the grep-only variant misclassified whitespaced/reserialized bodies as legacy. - telemetry fail-closed test now shadows bun with a failing stub (deterministic on any host layout); stale "five status cases" describe title renamed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: adversarial review fixes (Claude + Codex cross-model passes) Both adversarial passes ran against the wave; every FIXABLE finding landed with a regression test: - probeTimeoutMs clamps to >=1ms: a fractional override floored to 0, and execFileSync treats timeout:0 as NO timeout — the probe that exists to bound hangs could hang forever (found by both models independently). - /ship silent hook install now requires the hooks dir to live inside .git: with core.hooksPath (husky's COMMITTED .husky/), the chaining installer would have renamed the team's committed pre-push and written a machine-local wrapper into the working tree (found by both models). - gstack-config gbrain-refresh accepts the "timeout" status — the last consumer still gating on literal "ok" (Codex); gstack-gbrain-detect's config-derived fields honor GBRAIN_HOME so the detection JSON can't report status ok alongside config_exists false (Codex). - prepush: a remote sha absent locally (shallow clone / stale fetch) falls back to the merge-base/empty-tree range — scans MORE, never blocks a legitimate push into training users toward --no-verify. - dashboards: curl's own 000 no longer doubles to "HTTP 000000"; the community dashboard flags stale snapshots like the security one; array sections parse via jq (the sed/grep loops truncated at the first ']'); the no-jq marker grep tolerates whitespace. - telemetry: multi-line redactor output nulls the field instead of corrupting the JSONL record; setup's hint fires only when the config key is genuinely unset (an explicit false is a recorded decline); the /ship prompt marker honors GSTACK_HOME. Kept as designed (cross-model tension noted): Bearer stays MEDIUM in the prepush gate — a HIGH Bearer would block every docs example; the entropy validator can't eliminate that FP class, and MEDIUM warns visibly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: bump version and changelog (v1.57.11.0) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: P1 TODO — eval harness live progress + incremental persistence Root-caused during this ship: a killed eval run was indistinguishable from a healthy one for hours (per-file output buffering across mega test files, no incremental eval-store writes, no honest liveness signal). Full context and starting points in the entry. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test: fix operational-learning E2E fixture — copy lib/jsonl-store.ts Pre-existing breakage, proven on main: gstack-learnings-log has imported lib/jsonl-store.ts (shared injection patterns) since v1.57.5.0 / #1910, but the fixture copies only the bin scripts — the bin exits 1 before writing anything, on main silently (stderr swallowed) and on this branch loudly (the #1950 error-surfacing made the four-day-old failure visible). A real install always ships bin/ and lib/ together; the fixture now does too. Verified: the fixture-shaped invocation writes the learning (exit 0) with lib present, exits 1 on both main and this branch without it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ios-qa): isolate E2E tests under --concurrent (3 real races) The ios-qa E2E file failed intermittently under `bun test --concurrent` (the eval harness default). Three distinct shared-state races, all fixed: 1. Shared pidfile: a module-level `workDir` reassigned in beforeEach was clobbered by parallel tests, so concurrent daemons collided on the same pidfile and the loser returned `already_running`. Each test now gets its own dir via makeWorkDir(). 2. process.env path globals: tests set GSTACK_IOS_AUDIT_PATH / _ATTEMPTS_PATH / _ALLOWLIST_PATH on the shared process env; concurrent tests stomped each other's audit/attempts destinations. Threaded auditPath/attemptsPath/allowlistPath through DaemonOptions (and mintForCaller) as explicit args — env is no longer load-bearing. 3. afterEach cleanup race: the per-test cleanup drained a shared dir array, so the first test to finish deleted still-running tests' workDirs mid-assertion. Moved to afterAll (cleans once, after all settle). Verified: 5/5 clean full-suite runs at --max-concurrency 15 (was intermittent); daemon unit suite 91/91; daemon source compiles. The paths default to the env-derived locations when options are omitted, so the production CLI path is unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(pty): pin spawned claude to EVALS model chain (default claude-sonnet-4-6) launchClaudePty spawned the interactive `claude` TUI with no --model flag, so the child inherited the operator's ~/.claude/settings.json model. On a slow-thinking model that meant 5+ min of extended thinking on empty plan-mode context, timing out the plan-mode smoke tests regardless of contention. Pin the model via opts.model ?? EVALS_MODEL ?? 'claude-sonnet-4-6' — byte-identical to session-runner.ts:144, so PTY and `claude -p` evals always agree. Pushed before extraArgs (last flag wins, so a per-test --model still overrides). Placement leaves the spawn region byte-stable for a clean merge with the in-flight hermetic-env branch. Plumbed model through the three plan-skill wrappers. Static-grep tripwires guard the pin, its fallback chain, the before-extraArgs ordering, and all three wrapper forwards. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * test(pty): detect markdown bold-bullet prose AUQs (fixes office-hours smoke) office-hours auto-mode renders its mode question as `- **Building a startup**` markdown bullets (office-hours/SKILL.md.tmpl:102) with no letter/number marker. isProseAUQVisible only matched `A)`-style lettered or `1.`-style numbered options, so the question went undetected: the model surfaced it at ~2m19s (well under the 300s budget) but the harness kept scoring the run "working" off the spinner glyphs and timed out — a false timeout on a question that was already on screen. Add Pattern 3: when an interrogative line ('?') is present AND 3+ bold-bullet markers (`- **`) appear in the 4KB tail, classify as a prose AUQ. Bold is the discriminator vs incidental prose bullets; the line anchor is dropped (stripAnsi can collapse option lines) and the existing `❯ 1.` cursor gate still defers to a live native list. Wires through the existing classifyVisible 'asked' path and the timeout high-water-mark, so office-hours now classifies 'asked' instead of 'timeout'. Five unit cases: the office-hours render passes; no-'?', <3-bullet, plain-bullet, and native-cursor cases stay false. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * test(pty): detect stripAnsi-collapsed prose AUQs + judge spinner-precedence The plan-eng/plan-design plan-mode + finding-floor smokes timed out even when the skill HAD rendered a complete prose AskUserQuestion and was waiting: the PTY strips cursor-positioning escapes, collapsing the option newlines/spaces so "A) ..." arrives as "A(recommended)" / "-B:" and "Reply with A, B, or C" as "ReplywithA,B,orC". Every line-anchored detector (Patterns 1-3) returns false on those bytes, so proseAUQEverObserved never latched and the run timed out on a question that was already on screen. Add Pattern 4/5: a two-signal collapsed-form detector — a reply/recommendation marker (space-insensitive "reply with [A-D]", "Recommendation:", or "(recommended)") AND 2+ distinct A-D letters each punctuated by ) : or (. The conjunction is what separates a real AUQ from incidental report prose; verified true on the verbatim failing-run buffers where Patterns 1-3 return false. Also fix the Haiku judge spinner bias: of 614 verdicts, 569 were 'working' and 95 of those noted a question was visible — Claude Code keeps the spinner animating at an idle prose decision, so the judge coin-flipped. Add a precedence override: when an option list AND a Recommendation/Reply instruction are both visible, classify WAITING even with spinner glyphs. Kept the strict dual-signal gate (never option-list-alone) so auto-decide-preserved doesn't flip. 5 unit tests pin the two-signal contract (2 true on real collapsed bytes, 3 false guards). 90 -> 95 pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(plan-review): ask-first scope gate for plan-eng + plan-design review On an empty/cold invocation, plan-eng-review and plan-design-review would dive straight into repo exploration (plan-eng) or a 7-pass mockup+audit (plan-design) and only ask the user much later, if at all. plan-ceo-review already asks first via an unconditional Step-0 gate and behaves well; these two did not. Add a hard-STOP scope gate as the FIRST operational instruction in each skill (above the design-doc check / pre-review audit / mockup defaults it explicitly overrides): the first tool call must be AskUserQuestion confirming the review target, before any git/Read/Grep/Glob/Bash or mockup generation. Under --disallowedTools the options render as plain column-0 lettered prose with a Recommendation + "Reply with A, B, or C" line so the answer is detectable. This is correct cold-start UX (confirm what to review before grinding a full review on nothing) and it is the product half of the plan-mode smoke fix; the harness collapsed-form detector is the deterministic half that catches the ask however it renders. Templates + regenerated SKILL.md (default variant). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * test(tiers): reclassify stochastic plan-eng/plan-design ask-first smokes as periodic plan-eng-review and plan-design-review run a long explore/audit before their first AskUserQuestion, so whether the plan-mode + finding-floor smokes reach a terminal outcome within the 300s/600s budget depends on stochastic ask-first compliance (measured ~50-67%/run even with the hardened gate). Per the "non-deterministic -> periodic" tiering rule, move the four affected smokes (plan-eng/plan-design review-plan-mode + finding-floor) to periodic. The deterministic harness fix (collapsed-form detector + judge precedence) and the ask-first gate lift these from always-failing to mostly-passing and are the real product+harness improvements; periodic monitoring tracks the rate weekly without blocking PRs on an LLM coin-flip. plan-ceo/plan-devex ask-first reliably and stay gate-tier. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci(evals): gate the deterministic PTY plan-mode smokes in CI The real-PTY plan-mode smokes never ran in CI — the gate was local-only. Add an e2e-pty-plan-smoke matrix suite running the two deterministically-reliable ones (office-hours-auto-mode, plan-mode-no-op) so a regression there blocks PRs. The stochastic plan-eng/plan-design ask-first smokes stay periodic (touchfiles E2E_TIERS) and are not CI-gated. A fresh CI container has no ~/.claude.json, so the spawned interactive `claude` would wedge on the onboarding + API-key-approval dialog. Add a scoped seed step (hasCompletedOnboarding + key approval, its own ANTHROPIC_API_KEY env) before the run — mirrors what the hermetic E2E child env seeds. Per-suite timeout override (35 min) via matrix.suite.timeout so the PTY suite has headroom for --retry 2 without bumping the other 12 suites. Report runner count 12 -> 13. Validate via workflow_dispatch before relying on the gate (PTY-in-CI is new). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * ci(evals): install gstack skill registry for the PTY smoke suite The first dry-run of e2e-pty-plan-smoke failed: the spawned interactive `claude` printed "Unknown command: /plan-ceo-review". .claude/skills is gitignored, so a fresh CI checkout has no gstack skill registry and the TUI can't resolve /office-hours or /plan-ceo-review. Add a Register step (scoped to the suite, after Seed, before Run) that mirrors setup's --no-prefix user-scoped registry minimally: $HOME/.claude/skills/gstack -> repo (resolves the preambles' absolute ~/.claude/skills/gstack/bin/* and <skill>/sections/* paths) + per-skill SKILL.md/sections symlinks for the two skills these tests invoke. HOME is /github/home in this container and the runner adds no HOME/CLAUDE_CONFIG_DIR override (no hermetic mode), so $HOME is the right anchor — the Seed step already proved claude reads it. No ./setup (binary build + Chromium + fonts + /dev/tty prompt); SKILL.md + bin/ + sections/ are committed. Self-validating: fails the step loudly on a dangling symlink or missing `name:` frontmatter, so a moved target surfaces here instead of as a silent 35-min "Unknown command" timeout. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore: bump version and changelog (v1.58.4.0) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
|
|
Garry Tan
|
97584f9a59
|
feat(security): ML prompt injection defense for sidebar (v1.4.0.0) (#1089)
* chore(deps): add @huggingface/transformers for prompt injection classifier
Dependency needed for the ML prompt injection defense layer coming in the
follow-up commits. @huggingface/transformers will host the TestSavantAI
BERT-small classifier that scans tool outputs for indirect prompt injection.
Note: this dep only runs in non-compiled bun contexts (sidebar-agent.ts).
The compiled browse binary cannot load it because transformers.js v4 requires
onnxruntime-node (native module, fails to dlopen from bun compile's temp
extract dir). See docs/designs/ML_PROMPT_INJECTION_KILLER.md for the full
architectural decision.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): add security.ts foundation for prompt injection defense
Establishes the module structure for the L5 canary and L6 verdict aggregation
layers. Pure-string operations only — safe to import from the compiled browse
binary.
Includes:
* THRESHOLDS constants (BLOCK 0.85 / WARN 0.60 / LOG_ONLY 0.40), calibrated
against BrowseSafe-Bench smoke + developer content benign corpus.
* combineVerdict() implementing the ensemble rule: BLOCK only when the ML
content classifier AND the transcript classifier both score >= WARN.
Single-layer high confidence degrades to WARN to prevent any one
classifier's false-positives from killing sessions (Stack Overflow
instruction-writing-style FPs at 0.99 on TestSavantAI alone).
* generateCanary / injectCanary / checkCanaryInStructure — session-scoped
secret token, recursively scans tool arguments, URLs, file writes, and
nested objects per the plan's all-channel coverage decision.
* logAttempt with 10MB rotation (keeps 5 generations). Salted SHA-256 hash,
per-device salt at ~/.gstack/security/device-salt (0600).
* Cross-process session state at ~/.gstack/security/session-state.json
(atomic temp+rename). Required because server.ts (compiled) and
sidebar-agent.ts (non-compiled) are separate processes.
* getStatus() for shield icon rendering via /health.
ML classifier code will live in a separate module (security-classifier.ts)
loaded only by sidebar-agent.ts — compiled browse binary cannot load the
native ONNX runtime.
Plan: ~/.gstack/projects/garrytan-gstack/ceo-plans/2026-04-19-prompt-injection-guard.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): wire canary injection into sidebar spawnClaude
Every sidebar message now gets a fresh CANARY-XXXXXXXXXXXX token embedded
in the system prompt with an instruction for Claude to never output it on
any channel. The token flows through the queue entry so sidebar-agent.ts
can check every outbound operation for leaks.
If Claude echoes the canary into any outbound channel (text stream, tool
arguments, URLs, file write paths), the sidebar-agent terminates the
session and the user sees the approved canary leak banner.
This operation is pure string manipulation — safe in the compiled browse
binary. The actual output-stream check (which also has to be safe in
compiled contexts) lives in sidebar-agent.ts (next commit).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(security): make sidebar-agent destructure check regex-tolerant
The test asserted the exact string `const { prompt, args, stateFile, cwd, tabId } = queueEntry`
which breaks whenever security or other extensions add fields (canary, pageUrl,
etc.). Switch to a regex that requires the core fields in order but tolerates
additional fields in between. Preserves the test's intent (args come from the
queue entry, not rebuilt) while allowing the destructure to grow.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): canary leak check across all outbound channels
The sidebar-agent now scans every Claude stream event for the session's
canary token before relaying any data to the sidepanel. Channels covered
(per CEO review cross-model tension #2):
* Assistant text blocks
* Assistant text_delta streaming
* tool_use arguments (recursively, via checkCanaryInStructure — catches
URLs, commands, file paths nested at any depth)
* tool_use content_block_start
* tool_input_delta partial JSON
* Final result payload
If the canary leaks on any channel, onCanaryLeaked() fires once per session:
1. logAttempt() writes the event to ~/.gstack/security/attempts.jsonl
with the canary's salted hash (never the payload content).
2. sends a `security_event` to the sidepanel so it can render the approved
canary-leak banner (variant A mockup — ceo-plan 2026-04-19).
3. sends an `agent_error` for backward-compat with existing error surfaces.
4. SIGTERM's the claude subprocess (SIGKILL after 2s if still alive).
The leaked content itself is never relayed to the sidepanel — the event is
dropped at the boundary. Canary detection is pure-string substring match,
so this all runs safely in the sidebar-agent (non-compiled bun) context.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): add security-classifier.ts with TestSavantAI + Haiku
This module holds the ML classifier code that the compiled browse binary
cannot link (onnxruntime-node native dylib doesn't load from Bun compile's
temp extract dir — see CEO plan §"Pre-Impl Gate 1 Outcome"). It's imported
ONLY by sidebar-agent.ts, which runs as a non-compiled bun script.
Two layers:
L4 testsavant_content — TestSavantAI BERT-small ONNX classifier. First call
triggers a one-time 112MB model download to ~/.gstack/models/testsavant-small/
(files staged into the onnx/ layout transformers.js v4 expects). Classifies
page snapshots and tool outputs for indirect prompt injection + jailbreak
attempts. On benign-corpus dry-run: Wikipedia/HN/Reddit/tech-blog all score
SAFE 0.98+, attack text scores INJECTION 0.99+, Stack Overflow
instruction-writing now scores SAFE 0.98 on the shorter form (was 0.99
INJECTION on the longer form — instruction-density threshold). Ensemble
combiner downgrades single-layer high to WARN to cover this case.
L4b transcript_classifier — Claude Haiku reasoning-blind pre-tool-call scan.
Sees only {user_message, last 3 tool_calls}, never Claude's chain-of-thought
or tool results (those are how self-persuasion attacks leak). 2000ms hard
timeout. Fail-open on any subprocess failure so sidebar stays functional.
Gated by shouldRunTranscriptCheck() — only runs when another layer already
fired at >= LOG_ONLY, saving ~70% of Haiku spend.
Both layers degrade gracefully: load/spawn failures set status to 'degraded'
and return confidence=0. Shield icon reflects this via getClassifierStatus()
which security.ts's getStatus() composes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): wire TestSavantAI + ensemble into sidebar-agent pre-spawn scan
The sidebar-agent now runs a ML security check on the user message BEFORE
spawning claude. If the content classifier and (gated) transcript classifier
ensemble returns BLOCK, the session is refused with a security_event +
agent_error — the sidepanel renders the approved banner.
Two pieces:
1. On agent startup, loadTestsavant() warms the classifier in the background.
First run triggers a 112MB model download from HuggingFace (~30s on
average broadband). Non-blocking — sidebar stays functional during
cold-start, shield just reports 'off' until warmed.
2. preSpawnSecurityCheck() runs the ensemble against the user message:
- L4 (testsavant_content) always runs
- L4b (transcript_classifier via Haiku) runs only if L4 flagged at
>= LOG_ONLY — plan §E1 gating optimization, saves ~70% of Haiku spend
combineVerdict() applies the BLOCK-requires-both-layers rule, which
downgrades any single-layer high confidence to WARN. Stack Overflow-style
instruction-heavy writing false-positives on TestSavantAI alone are
caught by this degrade — Haiku corrects them when called.
Fail-open everywhere: any subprocess/load/inference error returns confidence=0
so the sidebar keeps working on architectural controls alone. Shield icon
reflects degraded state via getClassifierStatus().
BLOCK path emits both:
- security_event {verdict, reason, layer, confidence, domain} (for the
approved canary-leak banner UX mockup — variant A)
- agent_error "Session blocked — prompt injection detected..."
(backward-compat with existing error surface)
Regression test suite still passes (12/12 sidebar-security tests).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(security): add security.ts unit tests (25 tests, 62 assertions)
Covers the pure-string operations that must behave deterministically in both
compiled and source-mode bun contexts:
* THRESHOLDS ordering invariant (BLOCK > WARN > LOG_ONLY > 0)
* combineVerdict ensemble rule — THE critical path:
- Empty signals → safe
- Canary leak always blocks (regardless of ML signals)
- Both ML layers >= WARN → BLOCK (ensemble_agreement)
- Single layer >= BLOCK → WARN (single_layer_high) — the Stack Overflow
FP mitigation that prevents one classifier killing sessions alone
- Max-across-duplicates when multiple signals reference the same layer
* Canary generation + injection + recursive checking:
- Unique CANARY-XXXXXXXXXXXX tokens (>= 48 bits entropy)
- Recursive structure scan for tool_use inputs, nested URLs, commands
- Null / primitive handling doesn't throw
* Payload hashing (salted sha256) — deterministic per-device, differs across
payloads, 64-char hex shape
* logAttempt writes to ~/.gstack/security/attempts.jsonl
* writeSessionState + readSessionState round-trip (cross-process)
* getStatus returns valid SecurityStatus shape
* extractDomain returns hostname only, empty string on bad input
All 25 tests pass in 18ms — no ML, no network, no subprocess spawning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): expose security status on /health for shield icon
The /health endpoint now returns a `security` field with the classifier
status, suitable for driving the sidepanel shield icon:
{
status: 'protected' | 'degraded' | 'inactive',
layers: { testsavant, transcript, canary },
lastUpdated: ISO8601
}
Backend plumbing:
* server.ts imports getStatus from security.ts (pure-string, safe in
compiled binary) and includes it in the /health response.
* sidebar-agent.ts writes ~/.gstack/security/session-state.json when the
classifier warmup completes (success OR failure). This is the cross-
process handoff — server.ts reads the state file via getStatus() to
surface the result to the sidepanel.
The sidepanel rendering (SVG shield icon + color states + tooltip) is a
follow-up commit in the extension/ code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(security): document the sidebar security stack in CLAUDE.md
Adds a security section to the Browser interaction block. Covers:
* Layered defense table showing which modules live where (content-security.ts
in both contexts vs security-classifier.ts only in sidebar-agent) and why
the split exists (onnxruntime-node incompatibility with compiled Bun)
* Threshold constants (0.85 / 0.60 / 0.40) and the ensemble rule that
prevents single-classifier false-positives (the Stack Overflow FP story)
* Env knobs — GSTACK_SECURITY_OFF kill switch, cache paths, salt file,
attack log rotation, session state file
This is the "before you modify the security stack, read this" doc. It lives
next to the existing Sidebar architecture note that points at
SIDEBAR_MESSAGE_FLOW.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(todos): mark ML classifier v1 in-progress + file v2 follow-ups
Reframes the P0 item to reflect v1 scope (branch 2 architecture, TestSavantAI
pivot, what shipped) and splits v2 work into discrete TODOs:
* Shield icon + canary leak banner UI (P0, blocks v1 user-facing completion)
* Attack telemetry via gstack-telemetry-log (P1)
* Full BrowseSafe-Bench at gate tier (P2)
* Cross-user aggregate attack dashboard (P2)
* DeBERTa-v3 as third signal in ensemble (P2)
* Read/Glob/Grep ingress coverage (P2, flagged by Codex review)
* Adversarial + integration + smoke-bench test suites (P1)
* Bun-native 5ms inference (P3 research)
Each TODO carries What / Why / Context / Effort / Priority / Depends-on so
it's actionable by someone picking it up cold.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(telemetry): add attack_attempt event type to gstack-telemetry-log
Extends the existing telemetry pipe with 5 new flags needed for prompt
injection attack reporting:
--url-domain hostname only (never path, never query)
--payload-hash salted sha256 hex (opaque — no payload content ever)
--confidence 0-1 (awk-validated + clamped; malformed → null)
--layer testsavant_content | transcript_classifier | aria_regex | canary
--verdict block | warn | log_only
Backward compatibility:
* Existing skill_run events still work — all new fields default to null
* Event schema is a superset of the old one; downstream edge function can
filter by event_type
No new auth, no new SDK, no new Supabase migration. The same tier gating
(community → upload, anonymous → local only, off → no-op) and the same
sync daemon carry the attack events. This is the "E6 RESOLVED" path from
the CEO plan — riding the existing pipe instead of spinning up parallel infra.
Verified end-to-end:
* attack_attempt event with all fields emits correctly to skill-usage.jsonl
* skill_run event with no security flags still works (backward compat)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): wire logAttempt to gstack-telemetry-log (fire-and-forget)
Every local attempt.jsonl write now also triggers a subprocess call to
gstack-telemetry-log with the attack_attempt event type. The binary handles
tier gating internally (community → Supabase upload, anonymous → local
JSONL only, off → no-op), so security.ts doesn't need to re-check.
Binary resolution follows the skill preamble pattern — never relies on PATH,
which breaks in compiled-binary contexts:
1. ~/.claude/skills/gstack/bin/gstack-telemetry-log (global install)
2. .claude/skills/gstack/bin/gstack-telemetry-log (symlinked dev)
3. bin/gstack-telemetry-log (in-repo dev)
Fire-and-forget:
* spawn with stdio: 'ignore', detached: true, unref()
* .on('error') swallows failures
* Missing binary is non-fatal — local attempts.jsonl still gives audit trail
Never throws. Never blocks. Existing 37 security tests pass unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ui): add security banner markup + styles (approved variant A)
HTML + CSS for the canary leak / ML block banner. Structure matches the
approved mockup from /plan-design-review 2026-04-19 (variant A — centered
alert-heavy):
* Red alert-circle SVG icon (no stock shield, intentional — matches the
"serious but not scary" tone the review chose)
* "Session terminated" Satoshi Bold 18px red headline
* "— prompt injection detected from {domain}" DM Sans zinc subtitle
* Expandable "What happened" chevron button (aria-expanded/aria-controls)
* Layer list rendered in JetBrains Mono with amber tabular-nums scores
* Close X in top-right, 28px hit area, focus-visible amber outline
Enter animation: slide-down 8px + fade, 250ms, cubic-bezier(0.16,1,0.3,1) —
matches DESIGN.md motion spec. Respects `role="alert"` + `aria-live="assertive"`
so screen readers announce on appearance. Escape-to-dismiss hook is in the
JS follow-up commit.
Design tokens all via CSS variables (--error, --amber-400, --amber-500,
--zinc-*, --font-display, --font-mono, --radius-*) — already established in
the stylesheet. No new color constants introduced.
JS wiring lands in the next commit so this diff stays focused on
presentation layer only.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ui): wire security banner to security_event + interactivity
Adds showSecurityBanner() and hideSecurityBanner() plus the addChatEntry
routing for entry.type === 'security_event'. When the sidebar-agent emits
a security_event (canary leak or ML BLOCK), the banner renders with:
* Title ("Session terminated")
* Subtitle with {domain} if present, otherwise generic
* Expandable layer list — each row: SECURITY_LAYER_LABELS[layer] +
confidence.toFixed(2) in mono. Readable + auditable — user can see
which layer fired at what score
Interactivity, wired once on DOMContentLoaded:
* Close X → hideSecurityBanner()
* Expand/collapse "What happened" → toggles details + aria-expanded +
chevron rotation (200ms css transition already in place)
* Escape key dismisses while banner is visible (a11y)
No shield icon yet — that's a separate commit that will consume the
`security` field now returned by /health.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ui): add security shield icon in sidepanel header (3 states)
Small "SEC" badge in the top-right of the sidepanel that reflects the
security module's current state. Three states drive color:
protected green — all layers ok (TestSavantAI + transcript + canary)
degraded amber — one+ ML layer offline but canary + arch controls active
inactive red — security module crashed, arch controls only
Consumes /health.security (surfaced in commit
|