The gate treated a call as direct whenever only the recording proxy or a
custom tool's own proxy was in the path, so it stored IP pins the proxy
did not enforce. ProxyAddr now covers the recording and global proxies,
a per-tool check covers a custom http tool's own proxy, and the gate
marks the pin name-only for any of them.
Closes#368.
GET /api/conversations/{id}/messages passed limit through to the
database with no upper bound, so a huge query could load the whole
transcript. Clamp history pages to 200 and keep the since tail at 1000.
The stock seed was npx @playwright/mcp, but the UI mock and common
invocations use npx -y, which the first pass left untouched. Also add
--browser chromium to a leftover playwright-mcp --headless row, resolve
the MCP's playwright-core CLI via require.resolve, and cover proxy-arg
stripping so capture sync does not depend on args[0].
npx @playwright/mcp queries the npm registry on every start even with
the package installed globally, so the image's pin did not decide what
ran and offline starts failed. Seed playwright-mcp --headless --browser
chromium, rewrite leftover stock npx rows once, and install the Chromium
build the MCP's own playwright-core expects.
Fixes#329
POST /api/auth/change-password checked the current password with bcrypt
and had no limiter. A stolen session could guess without backoff.
Reuse the login limiter, keyed by client IP and by the session token,
so rotating the source address with the same token still blocks.
A successful change clears both buckets.
image_path was opened as a raw filesystem path, so any readable
png/jpg/jpeg/webp/gif on the host could be copied into the operator
transcript. The path is now resolved from the task working directory
and refused when it contains .. or EvalSymlinks outside that tree.
json.Unmarshal of JSON null into a map yields nil. syncBrowserMCPProxy
then assigned NODE_EXTRA_CA_CERTS under traffic capture and panicked,
including during NewManager on every restart once that row was stored.
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>
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.
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.
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.
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.
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.
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.
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.
The outbox migration is 0006 because 0005 is already traffic rewrite rules. A shutdown cancel leaves an in-flight intent claim running, so the next start can reset it to open instead of recording the interrupt as done or blocked.
Planner, worker, judge, and compactor rows rebuild the cap. Main-agent chat, goal decomposition, and side questions stay out, so a restart does not count turns the running process never metered.
The plaintext-substring guard encoded one encoder's output shape: it only held
while a small input compressed to fewer bytes than it started with, and Go
1.27 stores instead. Comparing sizes states the same invariant without
assuming how flate picks between stored and compressed blocks: a stored block
is larger than its input by construction, so "output is under half the input"
is true for any working encoder on a repetitive body and false for a body that
was never compressed.
Checked against the old 60-byte fixture on go1.27.4: it fails with "fixture is
not compressed: 60 bytes in, 85 bytes out", so the guard still catches the
case it was written for.