Commit Graph
1006 Commits
Author SHA1 Message Date
BaaRai 9b8ddf0610 Merge pull request #338 from Dev-next-gen/fix/llmpool-trimerr-utf8
fix(llmpool): keep trimmed provider errors on a rune boundary
2026-11-18 17:07:57 +01:00
Leo Camus 864a60d48b fix(llmpool): keep trimmed provider errors on a rune boundary
trimErr cut the error text at byte 300, which can split a multi-byte
character. The result is stored as llm_profile_health.last_error, and
PostgreSQL rejects the invalid UTF-8, so the breaker state of a profile
whose provider answers in Chinese is never persisted and the write is
retried every second.
2026-11-18 13:07:45 +00:00
BaaRai 230ba39dd5 Merge pull request #324 from cyberd4ch/docs/317-scope-distance-trusted-input
docs(security): document scope_distance as a trusted input
2026-11-18 12:26:26 +01:00
BaaRai 9767a70205 docs(security): name the real ingest counters and block rules
in_scope_not_asset counts in-scope events that are not asset types, not
gate rejects: a rejected candidate comes back as an insertion forbidden
error in the insert_assets result. EvaluateAssetGate also applies block
rules whether or not an allow-list exists, so the no-allow-list gap is
about the allow rule, not the whole gate.
2026-11-18 12:12:50 +01:00
BaaRai 366e14066f Merge pull request #322 from Dev-next-gen/fix/finding-filename-utf8
fix(report): keep finding zip names on a rune boundary
2026-11-18 12:12:23 +01:00
BaaRai 4c486e1c01 Merge pull request #328 from Michael-JRead/fix/surface-provider-refusals
fix: report provider refusals, do not rerun them as idle turns
2026-11-18 12:12:19 +01:00
Mike ReadandClaude Opus 5.5 03e216d032 fix: report provider refusals, do not rerun them as idle turns
A turn the provider's safety system declines (Anthropic stop_reason
"refusal", OpenAI-compatible finish_reason "content_filter") has no text
and no tool calls. Norma passes the stop reason through and the harness
ends any tool-less turn with ReasonCompleted, so the run summary said the
model "ended this turn normally", and steerHooks.Stop took the empty turn
for an idle one and resent the conversation up to the empty-response retry
count.

The LLM recorder is the one layer that sees the stop reason. It now hands
each completed call's stop reason to an optional context sink
(llmrec.WithStopSink). captureRunSession installs a per-run tracker
(agent.WithStopTracker) on that sink. With it:

- steerHooks.Stop does not inject the idle-turn rerun after a refusal;
- terminalText reports "provider refusal (stop_reason ...)" and says the
  intent was not completed, instead of the completed wording.

Other terminal states and ordinary empty turns are unchanged.

Fixes #327

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-18 00:41:18 -06:00
dach-a 94be0148ce docs(security): document scope_distance as a trusted input
States the trust boundary discussed in #317:

- output.json's scope_distance is the integrated scanner's own verdict
  and is ingested as-is; in-scope events are not re-checked against
  ARTEX's rule set.
- The enforced boundary is the task allow-list: assets reach the graph
  (insert_assets) only through EvaluateAssetGate, and rejects are
  recorded as in_scope_not_asset.
- The no-allow-list case is an accepted gap, with operator guidance:
  configure an allow-list per task or review output.json before feeding
  it to ARTEX. Cross-checking ingested events against ARTEX's rule set
  would redefine existing scanner configurations, so it is a deliberate
  design question, not an accepted vulnerability.

Option 2 as picked on #317; documentation only.
2026-11-18 03:15:57 +00:00
Leo Camus 88aefa0fc3 fix(report): keep finding zip names on a rune boundary
FindingFilename cut the name at 120 bytes, which can split a multi-byte
character. archive/zip only sets the UTF-8 flag on valid names, so the
entry was written without it and unzip tools decoded the whole name as
CP437, garbling every non-ASCII character in it.
2026-11-18 02:37:26 +00:00
BaaRai 51e2c603b9 Merge pull request #282 from renaesop/codex/fix-record-fact-input-contract
fix: repair record_fact summary contracts and asset aliases
2026-11-18 03:30:48 +01:00
BaaRai 9355d83eea Merge pull request #314 from yeting1107/docs/post-0-6-0-changelog
docs: record the post-0.6.0 changes
2026-11-18 03:03:44 +01:00
BaaRai 9b850a090c docs: add the post-#305 changelog entries
The rebase onto current main keeps the original entries and adds the three
user-facing changes merged after its base: the finding report_state field,
the ordered LLM health persistence, and the framed intent summaries.
2026-11-18 02:52:25 +01:00
yeting1107 398a08814a docs: record the post-0.6.0 changes 2026-11-18 02:52:06 +01:00
Chubboy Ren c833130c28 fix: enforce exact record_fact fields and validate asset anchors 2026-11-18 01:51:41 +00:00
Chubboy Ren c0c568b31f Merge remote-tracking branch 'upstream/main' into codex/fix-record-fact-input-contract 2026-11-18 01:34:26 +00:00
BaaRai ac5e714eb6 Merge pull request #313 from BaaRai/fix/task-llm-health-badge
fix(llm): record breaker health on every task path
2026-11-18 01:42:49 +01:00
BaaRai 6568a8eb44 fix(llm): record breaker health on every task LLM path
The /system/llm/ badge read llm_profile_health, which only the failover
Pool wrote. A task pinned to an explicit chain, a bound agent with
failover off, and the global provider all call the provider directly, so
a dead backend stayed "Healthy".

One policy now lives in llmpool.Registry.RecordCall, used by the Pool and
by the task runtimes: a success clears the state, a failover-eligible
failure trips it, context cancellation and deterministic 400s do not, and
a stream that already delivered output is left alone. The selection
carries the profile whose breaker the call updates, and a provider
already wrapped in the Pool reports nothing to avoid double counting.

The LLM recording page also explains an empty list while the llm_record
toggle is off, after the settings call resolves.

Originally reported in #257.
2026-11-18 01:31:18 +01:00
BaaRai 9a2a9e61ef Merge pull request #308 from Coff0xc/feat/finding-report-state
feat(findings): report stage separate from triage status
2026-11-18 00:17:48 +01:00
Coff0xc 6a0b79d34f feat(findings): add report_state separate from triage 2026-11-18 00:07:50 +01:00
BaaRai 729f935a21 Merge pull request #311 from BaaRai/fix/agent-frame-intent-summaries
fix(agent): frame intent summaries as data (issue #124)
2026-11-17 23:55:50 +01:00
BaaRai 9777163bbf fix(agent): frame intent summaries as data
The wake lines interpolated a planner-written intent summary into the trusted sentence, and the overview kept every summary in the plain JSON, so an inherited summary crossed tasks as if this planner had written it (#124, F3). Frame the summary on the wake lines and move the open/running/recent-done summaries into one framed block; the structured intent fields stay plain.
2026-11-17 23:44:36 +01:00
BaaRai e80a33c098 Merge pull request #301 from Protonmatter/protonmatter/session-epoch
Reject sessions issued before a password change
2026-11-17 23:28:24 +01:00
protonmatter 95bd9c24f0 Merge commit 'eddf03ce61f0a506fede039ea8936709b84ade8e' into protonmatter/session-epoch
# Conflicts:
#	CHANGELOG.md
2026-11-17 16:56:29 -05:00
protonmatter 134562c3d9 Drop the unused database import and align the password constants. 2026-11-17 22:54:16 +01:00
protonmatter 7cac9c240b Retry the password version on login when startup could not read it. 2026-11-17 22:54:16 +01:00
protonmatter ebc54b8751 Bind sessions to a password version so revocation cannot tie inside one second.
The hash and version commit together, a failed read rejects tokens, and the reset script tells the operator to restart.
2026-11-17 22:54:16 +01:00
protonmatter 26a7bbec37 Reject sessions issued before a password change.
A failed password read no longer mints a setup token, and filling the login limiter no longer drops an active lockout.
2026-11-17 22:54:16 +01:00
BaaRai eddf03ce61 Merge pull request #304 from Protonmatter/protonmatter/llmpool-health-queue
Coalesce circuit-breaker health writes
2026-11-17 22:52:27 +01:00
BaaRai 9bc5856565 Merge pull request #302 from Protonmatter/protonmatter/budget-restore
Restore task spend from the lanes the live ledger counts
2026-11-17 22:52:23 +01:00
protonmatter c9d81a453f Align the Server literal so gofmt accepts it. 2026-11-17 22:39:53 +01:00
protonmatter b2a2f95e8c Release the circuit-breaker queue lock around the database write. 2026-11-17 22:39:53 +01:00
protonmatter 1fbf2263a1 Keep circuit-breaker writes in state order and flush them before the database closes.
A later update can no longer be queued behind an older one, a failed save is retried, and shutdown waits for the last cooling window.
2026-11-17 22:39:53 +01:00
protonmatter 847e32407a Coalesce circuit-breaker writes onto one queue.
A newer state for the same profile replaces one that has not been written yet, and a failed write is logged instead of discarded on a detached goroutine.
2026-11-17 22:39:50 +01:00
BaaRai 7bd9bf6e3e Merge pull request #309 from BaaRai/feat/traffic-rewrite-p33b1
feat(traffic): apply rewrite rules in the capture proxy (P3.3b1)
2026-11-17 22:38:59 +01:00
BaaRai f8bc3b3b6d Merge pull request #305 from Protonmatter/protonmatter/json-body-cap
Cap JSON request bodies
2026-11-17 22:29:08 +01:00
BaaRai 4fb29fd841 Merge pull request #303 from Protonmatter/protonmatter/coverage-flag-error
Keep coverage tools available when the flag read fails
2026-11-17 22:29:04 +01:00
protonmatter 074962a351 Say that a revived task keeps the spend already recorded. 2026-11-17 16:25:16 -05:00
BaaRai 7c3d50d533 docs: describe applying the rewrite rules in the proxy
The status line said the proxy does not apply rules yet. Replace it with
the request and response phases, what exchange_rewrites records and how
the hashes are formed (plain SHA-256, keyed HMAC for sensitive actions),
the SSE and WebSocket behaviour, the 5 MiB response pass-through, the
snapshot limits (live slots are never deleted by the TTL), and the fact
that replays go through the proxy and are rewritten like captures.
2026-11-17 22:23:12 +01:00
BaaRai b57bafe74a feat(server): upgrade the built-in traffic_get description once
It now says that traffic_get shows the originals of a rewritten exchange
and that traffic_search never finds a value a rewrite rule wrote. As for
traffic_search, only the original built-in text is replaced.
2026-11-17 22:23:12 +01:00
BaaRai fa07d69f1d feat(traffic): apply rewrite rules in the capture proxy (P3.3b1)
The sink runs the request phase in the Request hook and the response phase
at the start of the Response hook; at SSEStart only status and header
actions apply. WebSocketStart comes after the client is upgraded, so the
handshake is traced "not rewritten: stream" and recorded as sent.

Exchanges keep what the client sent and what the target answered. The
original of a rewritten request lives in a per-flow snapshot slot (entry or
tombstone, 1024 entries / 64 MiB, 4096 tombstones), reserved before the
request is changed, released on Response, RequestError, SSEStart, a 5 MiB
response, or by a sweep of finished flows (ClientDisconnected, 1 s
debounce, 1 min period). Under pressure an entry older than 15 minutes is
turned into a tombstone in place, only when that frees room for the new
reservation; its exchange is recorded "original request unavailable".

record(), recordStreamStart() and insertExchange() clear any trace under
the id before INSERT OR REPLACE; the trace (exchange_rewrites, created
with CREATE TABLE IF NOT EXISTS) is written in the exchange's transaction,
its spilled values registered in blob_refs, and purged with the exchange.
The full-text index covers the originals only. Logs of a rewritten flow
name the host, the step or error class and the rule ids, never the URL or
the error text. A rewritten request's transport error never switches its
host to passthrough. traffic_get shows the rewritten flag and rule ids.

Without a rule that applies, what is recorded is unchanged byte for byte.
2026-11-17 22:23:12 +01:00
BaaRai b3e1f5359f feat(traffic/rewrite): let the proxy record a rule's transport error
The request phase can fail after the engine is done: the rewritten request
does not reach the target. NoteError records that as the rule's last error,
under the new class transport_error.
2026-11-17 22:22:56 +01:00
BaaRai 5fdf8eb075 fix(server): keep the MaxBytesError so a caller's 413 mapping still works
A handler that installs its own MaxBytesReader and then calls decode lost the 413: decode converted the error to a plain one and the middleware's limitedBody had been replaced. Return the original *http.MaxBytesError and update the cap test to errors.As.
2026-11-17 22:19:59 +01:00
BaaRai d646f6999b fix(test): check each coverage tool on the set that carries it
list_untested_assets is a planner tool, not a worker tool; the fail-open assertion can only be made against PlannerTools.
2026-11-17 22:18:54 +01:00
BaaRai 8ae53f74cf Merge pull request #306 from Protonmatter/protonmatter/mitm-passthrough-ttl
Expire MITM passthrough after one minute
2026-11-17 22:14:58 +01:00
BaaRai 03ac1c7d47 Merge pull request #300 from Protonmatter/protonmatter/shutdown-outbox
Drain task goroutines on shutdown and replay unsaved activities
2026-11-17 22:08:25 +01:00
protonmatter 287cbb685c Treat an expired passthrough mark as intercepted again, and say what the window costs.
Replay was still honoring a stale bypass, and a host whose interception always fails now loses one request per minute instead of one per process.
2026-11-17 15:58:24 -05:00
protonmatter 186594ed3a Cap every non-multipart body and answer an oversized one with 413.
Handlers that decode JSON directly were still unbounded, and a body past the cap was reported as a bad request.
2026-11-17 15:57:43 -05:00
protonmatter 133dc68668 Test that a failed coverage read stays enabled, including auto-scope writes.
A missing task and a query error both leave the flag on. The worker keeps the coverage tools and still records auto scope when the stored flag is off.
2026-11-17 15:52:04 -05:00
protonmatter 40de6ba898 Say what a restored task budget still allows, and test the lanes that count.
A task already inside the wrap-up reserve only has that reserve left. The worker estimate starts from the defaults because usage rows have no run boundary.
2026-11-17 15:50:00 -05:00
protonmatter a1c6103cbf Stop admitting work before the drain and replay each outbox row once.
The listener closes before Wait, and a shutdown refuses new task operations. Wait also stays until detached runs finish. The activity insert and the outbox delete commit together, and a row that keeps failing is dropped after five attempts so it cannot block the rest.
2026-11-17 15:41:56 -05:00