* feat(a2a): add A2A v1.0 agent-interop channel plugin
Adds `a2a` as a bundled, default-off channel plugin so A2A-compliant agents
(Hermes Agent, LangChain, CrewAI, Google ADK) can discover an OpenClaw gateway
and send it tasks, and so OpenClaw can address configured peers.
Inbound HTTP (`/.well-known/agent-card.json`, `/a2a/v1`) runs through the normal
channel ingress path, so peer messages inherit allowlist admission, provenance
framing, and session routing. The agent reply returns through the channel
delivery callback and resolves the A2A task with its artifact, correlated per
(peer, contextId) FIFO so concurrent sends cannot cross-talk.
Wire format follows the canonical v1.0.0 spec sources (a2aproject/A2A
specification/a2a.proto): PascalCase SendMessage/GetTask/CancelTask,
supportedInterfaces[] rather than the 0.3-era top-level url/protocolVersion,
member-presence Part oneof, TASK_STATE_*/ROLE_* enums, no Task.kind. The 0.3
dotted method names are accepted as explicit compat aliases because shipped
Hermes-generation peers still send them.
Security: per-peer bearer tokens compared in constant time, per-peer sliding
window rate limit, 1 MiB body cap, 64 KiB inbound text cap, tasks scoped to
their owning peer, outbound redirect: "error", and no inbound-controlled target
URLs. With no peers configured the channel stays unconfigured and registers no
routes.
Live-verified against a real gateway on an isolated state dir with a mock
provider: 31/31 proofs, including the agent reply round-tripping into the task
artifact. That run caught two defects fixed here:
- The agent card read `cfg.agents.list` directly, so gateways configured with
the canonical `agents.entries` roster advertised zero skills. Now uses
listAgentIds/resolveAgentConfig, which read either roster shape.
- `returnImmediately` dispatches inherited the HTTP request's released work-
admission root and every async task failed as GatewayDrainingError. Now
reserves an independent root via runDetachedWebhookWork, matching sms/line/zalo.
* fix(a2a): route outbound sends through the SSRF guard and land setup metadata
CI on the first push surfaced four real defects that local `check:changed` did
not cover (it runs no oxlint or typecheck lane):
- Outbound peer sends used raw `fetch()`, tripping the channel/plugin
no-raw-fetch boundary. They now go through `fetchWithSsrFGuard` with
`maxRedirects: 0`, so A2A egress carries the same SSRF policy as every other
plugin call and a redirect cannot re-target a delivered task.
- The channel declared an empty setup contract while its package manifest
declared none, so the bundled-channel shape contract failed. The wizard now
collects `advertisedUrl` plus a first peer name/token pair, which is the
minimum that leaves A2A actually configured (it stays unconfigured until one
peer credential exists), and the package manifest mirrors that metadata.
- Six lint errors in the test files: base-to-string on `BodyInit` bodies, an
unused mock, and a shadowed `options` parameter. Request bodies now go through
a typed reader that asserts the serialized-string contract before parsing.
- A card assertion referenced `capabilities` without declaring it on the local
response type, failing test typecheck.
Outbound redirect handling moves from `redirect: "error"` to the guard's
`manual` inspection; the adapter tests assert the new shape.
* fix(a2a): drop route replaceExisting, accept configWrites, register runtime api
CI after the interop run surfaced four more gates:
- `registerPluginHttpRoute` used `replaceExisting: true`, tripping the
GHSA-RQP8-Q22P-5J9Q rule: a duplicate path can silently replace another
account's handler. A2A owns fixed global paths on a single account, so a
duplicate registration means a stale or conflicting owner. It now fails loudly
through `throwOnFailure` instead.
- The channel schema rejected `channels.a2a.configWrites`, which every bundled
channel must accept; added to the zod schema, the manifest schema, and the
config type, then regenerated the channel metadata.
- The new `input` assertion in the setup adapter needed a `// SAFETY:` line for
the assertion ratchet.
- `extensions/a2a/runtime-api.ts` needed registering in the runtime-api
classification list, like every other bundled channel barrel.
Also repairs a pre-existing incomplete `vi.mock` factory for `../infra/fs-safe.js`
in the agent-delete suites. Both pass in isolation, but the shared-worker lane
lets a sibling importer reach `readLocalFileSafely`, which the partial factory
never exported; adding this channel changed shard composition and surfaced it.
The factory now spreads the real module and overrides only `movePathToTrash`.
* chore(a2a): refresh config baseline for the configWrites surface
* chore(tests): adopt main's fs-safe mock repair for the agent-delete suites
Main landed the same importOriginal spread for the shared-worker mock-defeat
while this branch carried an equivalent fix; take main's version verbatim so the
two do not conflict.
* fix(a2a): isolate peer sessions and stop faking task cancellation
Addresses both ClawSweeper P1 findings.
Peer sessions: inbound routing never passed `dmScope`, so it fell back to
`session.dmScope ?? "main"` and every authenticated peer converged on
`agent:main:main` under default config. Untrusted remote content was joining the
operator's own session, and peers shared conversation history. A2A now pins
`per-account-channel-peer`; the peer id already embeds the A2A contextId, so each
peer+context pair gets its own session regardless of global session config.
Cancellation: `CancelTask` marked the task `TASK_STATE_CANCELED` and discarded
the late reply, but the dispatched agent run kept going and could still use
tools. The peer was told the work stopped when it had not. There is no
plugin-facing seam to abort a live run, so cancellation is refused with -32004
instead of acknowledged, and the store's cancel path plus its canceled-delivery
tombstone are deleted rather than left as a lying terminal state. Documented as
an explicit limitation.
Live-verified on a real gateway: 34/34 proofs, including three distinct
`agent:main:a2a:default:direct:<peer>:<context>` sessions with zero traffic in
`agent:main:main`, and both cancel spellings refused while the task stays
WORKING. Re-ran the official a2a-sdk 1.1.2 Docker interop afterwards: 9/9.
* ci(macos): budget macos-swift by runner so fork PRs finish
runs-on falls back to hosted macos-26 for several cases: workflow_dispatch,
run_attempt > 1, fork pull requests, and (as of the runner-backend circuit
breaker landed on main the same day this PR was authored) breaker-routed
GitHub-hosted fallback. timeout-minutes did not cover the fork-PR case, so
fork PRs ran on the slow hosted runner with the Blacksmith-tuned 20-minute
budget.
Measured on PR #118989 (a fork PR): macos-swift was cancelled after 20m25s,
killed mid-compile at step 1365/1416. No test executed, and the log has no
swift compile error - only 'The operation was canceled.' ci-gate then fails
because it lists macos-swift as required, so the PR reads red for a reason
unrelated to its diff, and a contributor cannot rerun it.
Make the budget follow the runner instead of the trigger: every hosted path
gets 30 (folded into the same budget the circuit breaker's own hosted-fallback
timeout extension already established for this job), so this doesn't
reintroduce a second, competing hosted-timeout value. Blacksmith paths keep
20 unchanged. Replaces the single pinned-string guard with a table-driven
test covering runs-on and timeout-minutes together across every trigger
context that can route to a hosted runner.
* chore: refresh PR head (keep open for maintainer review)
* ci(macos): track main's author-association runner routing
Main now routes macos-swift by pull_request author_association rather than
fork-ness, so the timeout predicate and its guard scenarios follow it.
* ci(macos): tighten hosted runner budget coverage
Co-authored-by: harjoth <harjoth.khara@gmail.com>
* docs(ci): document hosted macOS budgets without merge conflicts
Co-authored-by: harjoth <harjoth.khara@gmail.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(release): preserve VCR mirror source digests
Transport only attestation-verified digests across secret-scanned job outputs, reconstruct immutable GHCR refs inside the VCR mirror, and add an approved mirror-only recovery path.\n\nCloses #129466
* fix(release): verify VCR recovery sources
Revalidate attestations and release-version labels before any VCR registry write so manual recovery preserves the immutable source boundary.
* test(release): keep VCR regression scoped
Leave global workflow-to-test routing cleanup for a follow-up; this PR directly changes and runs both VCR regression suites without forcing metadata-complete CI.
* fix(ci): preserve caches after warmer failures
Finish every selected cache-warm group, save content-keyed transform and compile caches, then fail visibly after the save steps. Ordinary CI remains fail-fast.
Queue Telegram-visible Mantis proofs at workflow level so only one shared-user run allocates a runner at a time.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* feat(geolocation): resolve client addresses to a coarse city via a bundled plugin
The Activity identity card could show a client's IP address but not where it
was, so an operator still had to look the address up by hand.
Add a bundled `geolocation` plugin that owns address-to-place resolution behind
one authenticated route, `GET /plugins/geolocation/lookup?ip=`. It downloads a
MaxMind-format database on first lookup into the state directory, answers from
that local copy, and refreshes it monthly, so a lookup never sends an address
to a third party. The Control UI renders the resolved city on the device row
next to the address and the client-reported time zone.
The default source is DB-IP City Lite under CC BY 4.0. That license requires
attribution, so every response carries the credit and the UI renders it next to
the value; the database is downloaded at runtime and never redistributed.
Plugin code and the `maxmind` reader are MIT. No free city-level IP database is
MIT-licensed, so the obligation lives with the data rather than the code, and
`databaseUrl` plus the attribution fields make the source swappable.
No new core provider kind: with one implementation the plugin owns everything
through the existing HTTP-route seam, keeping core plugin-agnostic. A second
provider is what would justify promoting this to a registry contract.
Availability and lookup failure stay distinguishable: a missing or still
downloading database answers 503, never `found: false`. A failed refresh serves
the cached copy, and a body that does not parse as an MMDB is discarded without
replacing a working database.
* fix(docs): correct geolocation config examples and add zh-CN glossary entries
The config examples used `plugins.<id>` instead of the real
`plugins.entries.<id>.config` shape, which the docs config-example
validator and src/config/docs-config-examples.test.ts both reject.
New doc labels also need zh-CN glossary entries.
* chore(labeler): cover the geolocation extension directory
AGENTS.md requires a labeler entry plus a GitHub label for every new
plugin surface; test/scripts/labeler-extension-coverage.test.ts enforces
the labeler half.
* fix(geolocation): address review findings on caching, download bounds, and scope
Cold-start lookups were permanently suppressed. The loader cached one promise
per address including failures, so the 15s browser deadline expiring against a
first download that takes ~46s cached a blank forever, and a mounted row only
looks up again when its IP changes. Lookups now return a discriminated
located/absent/unavailable result: only definitive answers are cached, and the
element retries an unavailable one on a widening 5s/15s/45s schedule.
Download limits ran after allocation. The size check happened only after
`response.arrayBuffer()` had buffered the whole body, and gunzip had no output
ceiling, so a replaced source or a compression bomb could exhaust Gateway memory
before rejection. The body now streams against a compressed ceiling enforced
per chunk, and inflation uses zlib's maxOutputLength.
Cached placements were not scoped to the Gateway. The cache keyed only by
address while endpoint and credentials come from the shared Gateway context, so
a switch could render the previous Gateway's answer. The shared reset hook now
supports multiple subscribers - a single slot silently dropped whichever
registered first - and the geolocation cache subscribes.
Unresolvable ranges no longer trigger a download. Only loopback suppresses `ip`
at connect, so Tailscale carrier-grade-NAT and LAN addresses are recorded and
displayed. No geolocation database contains them, so a tailnet-only or LAN-only
Gateway was downloading 125 MB to answer nothing. The route now answers those
ranges without loading the database, using the already-public
`isPrivateOrLoopbackHost` seam so the SDK surface budget is unchanged.
The quickstart queried a reserved documentation range while showing a located
response, which cannot happen; it now uses a routable address and documents the
not-found case.
* fix(deps): resync the lockfile after dropping the net-policy dependency
The geolocation plugin briefly depended on @openclaw/net-policy before
switching to the already-public isPrivateOrLoopbackHost SDK seam. The
package.json entry was removed without regenerating the lockfile, so the
frozen-lockfile install failed and every downstream CI job failed with it.
Remove the bundled OpenProse plugin and /prose command now that upstream owns the maintained Agent Skill. Preserve /prose as migration documentation and let Doctor clean stale plugin configuration.
BREAKING CHANGE: The bundled OpenProse plugin and /prose command are removed.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix(scripts): size the tsdown heap from the build's own cgroup budget
The build heap probe only read the cgroup root (/sys/fs/cgroup/memory.max and
the v1 equivalent). Those files exist only when the process runs in a
namespaced container cgroup; under systemd the budget lives on the process's
own slice, and the v2 root carries no limit at all. So every systemd-managed
build found no limit, fell back to /proc/meminfo MemTotal, and took the full
12288 MB default heap regardless of its actual budget.
Observed on a 15.4 GiB host: openclaw-main-update.service ran tsdown with
NODE_OPTIONS=--max-old-space-size=12288 while its user@999.service slice was
bounded at 5 GiB, reaching 3.2 GB RSS and 6.25 GB peak before the host began
OOM-killing unrelated services.
Resolve the limit from /proc/self/cgroup and walk that chain instead, reading
memory.high alongside memory.max (memory.high throttles reclaim rather than
failing allocation, so a heap above it stalls the build instead of OOM-ing),
and take the tightest bound found. Root paths stay as the container fallback,
and an explicitly injected path list still disables detection.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): resolve the build heap budget from the v1 memory controller too
The slice walk only accepted the unified 0:: record, so a legacy or hybrid
systemd host fell back to the root probe and kept taking host memory. One
resolver now walks both hierarchies leaf-to-root, which makes the static root
list its own depth-0 case and removes it.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): read cgroup controller mounts instead of assuming their paths
v1 controllers can be co-mounted at the cgroup root, where memory.limit_in_bytes
sits under the slice with no per-controller directory, so the hardcoded
/sys/fs/cgroup/memory probe missed the budget and the build took the full
12288MB default. Mount points now come from mountinfo.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): translate cgroup records through the mount root
mountinfo field 4 is the subtree a cgroupfs mount exposes. Under a container
mount the /proc/self/cgroup record stays host-absolute, so walking it verbatim
probed paths below the visible mount and the build fell back to host memory.
Records now translate through the mount root before the walk.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): skip cgroup mounts that cannot represent this process
Falling back to the mount root for a record outside the mount's subtree sized
the build from an unrelated cgroup: an inherited namespace clamped the heap to
the 2048MB floor from a foreign 1GiB limit. Non-representable mounts are now
skipped, and the blind root probe only runs when no memory record exists.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): keep every cgroup mount view, not just the last one seen
One hierarchy can be visible through several mounts and only some expose a
subtree containing this process. Retaining only the last view dropped the
budget whenever a non-representable bind view came later, sending the build
back to host MemTotal.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): decode octal-escaped mountinfo paths before matching cgroups
ClawSweeper P2 on 7e64ad61f7: the cgroup resolver compared mountinfo's mount
root and mount point verbatim. The kernel escapes space, tab, newline, and
backslash in those two fields, so any cgroup mounted under such a path never
matched, the bounded slice was missed, and heap sizing silently fell back to
host memory.
Decode both fields before matching. The decoder lives in scripts/lib beside the
other shared script helpers rather than inline, so the scripts program has one
copy rather than a new ad hoc one.
Regression test fails pre-fix: a v2 mount at "/sys/fs/cgroup\040dir" with a
5 GiB memory.high yields --max-old-space-size=12288 (host fallback) before the
fix and 4352 after.
Follow-up, deliberately not bundled here: src/infra/sqlite-wal.ts,
src/commands/doctor-state-integrity.ts, and src/plugins/bundled-source-overlays.ts
each carry their own private copy of this same decoder. Consolidating all four
into @openclaw/normalization-core is the right end state, but it touches a
shared package plus three core modules and belongs in its own reviewable change.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): resolve cgroup-namespace-relative records to their mount
ClawSweeper P1 on d6fe49dd3f: inside a cgroup namespace /proc/self/cgroup
reports the namespace root ("0::/") while mountinfo field 4 stays the host
subtree the cgroupfs was mounted from ("/docker/<id>"). relativeCgroupPath then
found no prefix match and returned null; because a memory record had already
been seen, the root probe was skipped and the build fell back to host MemTotal.
A constrained container therefore missed its own budget entirely.
That namespace root is exactly what the mount exposes at its mount point, so it
resolves to "/" rather than failing closed.
Regression test fails pre-fix: a "0::/" record against a /docker/2f1a9c mount
root with a 5 GiB memory.max yields --max-old-space-size=12288 before the fix
and 4352 after.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): reject inherited cgroup mount views instead of guessing
ClawSweeper P1 on b4d200c5d2: the previous commit resolved a namespace-relative
record against any mount root, including the inherited views cgroup_namespaces(7)
documents, whose field-4 root reads "/..". Which cgroup such a view exposes is not
derivable from mountinfo, so probing it can size the build from an unrelated
cgroup's limit.
Reject non-canonical mount roots outright. An undecidable view now falls back to
host sizing, which is current main's behavior, rather than silently adopting the
wrong budget.
Regression test covers the "/.." inherited mount: it must yield host MemTotal
sizing, not the 5 GiB limit sitting behind that mount.
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(scripts): fail closed on namespace-root records against non-root mounts
ClawSweeper P1 on 731d3bbc8e: a "0::/" record does not prove that a mount
rooted at some other subtree exposes this process's cgroup. Resolving that pair
could cap the build heap from an unrelated cgroup's limit.
Return no mapping for it. An undecidable pair now falls back to host sizing,
which is current main's behavior, so the failure mode is a missed optimisation
rather than a wrong budget. The "/.." inherited-mount rejection stays; this
covers the broader ambiguous mapping it did not.
The namespace-relative test is repointed accordingly: an unrelated mounted
subtree must yield host sizing, not that subtree's limit.
Net production change: none (4 lines swapped).
Co-authored-by: jesse-merhi <79823012+jesse-merhi@users.noreply.github.com>
* fix(build): cap tsdown heap to the real budget and refuse hosts that cannot build
The 2048MB floor was applied on top of a discovered cgroup limit, so a small
container was handed a heap larger than it could honour. Measured in real
cgroups, that does not OOM-kill, it thrashes: a 1500MiB container sat pinned at
its ceiling for 10 minutes with oom_kill at 0, never finished the second of
eleven invocations, and starved every other process on the host.
Cap to the discovered budget, then refuse up front when that budget cannot hold
the build. The threshold is the whole-build peak, not a single pass: a full
eleven-invocation build peaks at 4730MiB, so a 5GiB slice completes while 4GiB
and 2816MiB slices are both killed partway through the third invocation.
The refusal runs before any output is cleaned, so a host that cannot rebuild
does not also lose the build it has.
* fix(build): harden tsdown heap admission
* fix(build): guard the default tsdown plan
* fix(build): preserve runtime-only Docker builds
* fix(build): admit only declaration cache misses
* fix(build): scope heap admission to real budgets
* fix(build): guard direct unified declarations
* fix(build): guard the canonical tsdown config
* fix(build): satisfy cache planning lint
* fix(gateway): release empty orphan leases
* fix(build): cap cgroup budget by host memory
* fix(build): serialize the canonical tsdown config
* test(build): freeze host memory fixtures
* fix(build): honor cgroup v1 soft limits
* fix(build): respect cgroup v1 hierarchy mode
* fix(build): admit unified runtime plans
* fix(build): admit every unified runtime path
* fix(build): collect repeated tsdown filters
* fix(build): ignore cgroup v1 soft limits
* fix(build): use explicit heap override as opt-in
* refactor(build): simplify memory admission
* fix(build): harden constrained build recovery
* fix(ci): prebuild runtime before real CLI shards
* fix(build): honor runtime-only runner environment
* fix(ci): satisfy tooling shard lint
* fix(mantis): enforce verdict-expectation coherence and publish agent analysis files
Mantis run 32619081130 on #127989 published an overall `pass` while its own
manifest recorded that the candidate expectation was not observed: per-lane
`status` was mechanical capture success and the agent's judgment lived only
in `expected` prose, so nothing reconciled the two before publication.
- `mantis-evidence.json` schemaVersion 2: each comparison lane carries a
required boolean `expectationMet`. The desktop agent sets it in the same
manifest edit as `expected`; mechanical producers (Telegram live, web UI,
Slack, Discord) derive it from lane status.
- `scripts/mantis/publish-pr-evidence.mjs` is the single enforcement owner:
it requires the booleans, recomputes `pass`/`outcome`, downgrades a
contradictory pass claim to `fail`, and renders a visible "verdict
downgraded" note. The desktop workflow invokes it with `--validate-only`
before upload or comment.
- Agent top-level `*.json`/`*.md` analysis files (assertions, comparisons,
recipe suggestion) now survive the quarantine rebuild and upload, so cited
evidence actually exists in the artifact.
* fix(mantis): derive expectations from trusted lane facts
Capture the proven mechanics from proof run 32616859410 (PR #127950) as a
reusable recipe: dual 600s timeout config patch, Code Mode exec events file,
streamed chunkDelayMs hold, and the observe-loop pattern around the 60s cap.
Require agents to record recipe-suggestion.md after scenario-mechanics
iteration so future discoveries persist.
* feat(mantis): add exec and restart lane primitives
Give the proof agent a developer shell inside each SUT container and an
in-container gateway restart so it can design scenarios like a local
developer: patch openclaw.json and restart, stage plugins and fixtures,
run node/tsx against the read-only repo root, inspect SQLite state.
- container script: exec (docker exec as mantis-sut, bounded by timeout),
restart (request file + TERM), sut_command becomes a relaunch supervisor
- lane CLI: exec returns bounded stdout/stderr/exitCode and records a
redacted invocation; restart waits for a fresh [gateway] ready marker
- MAX_SENDS 12 -> 40 (shared-QA-bot flood safety, not a scenario bound)
- runtime root chown root:mantis-proof, mode 1770 so the agent can stage
files while root-owned attestation stays unreplaceable
* docs(mantis): let the proof agent design scenarios like a local developer
Lead with developer-shell parity, allow reading whatever code the scenario
needs (PR text still untrusted, PR code only inside SUT lanes), document
exec/restart shapes, and reserve block for hard impossibilities.
* fix(mantis): keep the SUT exec result type local
* fix(mantis): resume the agent when it ends without a manifest
Run 32615428295 (exec branch on #127950) hit Codex context compaction at
03:52:11 and the model answered with a confabulated "handoff" message instead
of continuing; codex exec exited 0 with no mantis-evidence.json and the
trusted-evidence step failed the run with no verdict.
The agent step now checks for the manifest after codex exits and, when it is
missing, resumes the same thread (`codex exec ... resume --last -`, verified
against codex-rs/exec/src/lib.rs at rust-v0.149.0: cwd-matched latest thread,
`-` reads the prompt from stdin) with a short correction prompt, bounded to
three resumes. The main prompt states that a handoff/summary is never an
acceptable final message.
The proof gateways execute runtime JS only; declarations forced a ~177s
unified rebuild per lane because the declaration cache key is an
aggregate source hash. OPENCLAW_RUN_NODE_SKIP_DTS_BUILD=1 on profile
full now selects the runtime artifact surface via the uncached generic
tsdown graph, both Mantis lanes pass it through, and the baseline
archive moves to an isolated mantis-runtime-v1 namespace. Measured
candidate build: 177.1s -> 33.05s.
The Telegram Desktop proof previously built the candidate from the raw PR
head, so a head behind main failed for reasons main already fixed
(observed on #127770: 59 commits behind, hitting the Unknown-model defect
fixed by #127952). GitHub's cached PR test merge cannot be the candidate
either: it was observed 50 commits / 12 hours stale and never refreshed.
The workflow now resolves the live refs/heads/main tip and merge base at
dispatch, requires a main-targeting PR, and builds a deterministic local
merge (merge-tree --merge-base + commit-tree with pinned identity/date) as
the candidate; conflicts fail with a direct rebase message and there is no
fallback to the raw head. Lane labels and docs now say the candidate is
the PR merged onto main.
The pdf tool auto-resolves its model from plugin manifest defaults
(openai gpt-5.5), which the Mantis SUT catalog does not define, so every
pdf round trip died with "Unknown model: openai/gpt-5.5" before
native-vs-fallback dispatch (run 32572580656). Pin
agents.defaults.pdfModel to the only harness model and account for the
pdf tool's own provider request in the staged-media recipe wait/pins.
* docs(mantis): script catalog-tool turns through Code Mode exec
Run 32570733988 blocked both lanes with 'Tool pdf not found': the proof
agent scripted a top-level pdf function_call, but the SUT agent runs Code
Mode where catalog tools execute from JavaScript inside exec. Document the
canonical route in the proof prompt, add a complete exec(pdf) scripted-event
example plus the document-extract configPatch to the staged-media recipe,
and pin the new guidance in the workflow test.
* docs(mantis): wait for the exec round trip before finish
ClawSweeper P2: send releases the mock response once the outbound message
is visible, and finish tears the lane down, so a slow exec/pdf round trip
could lose its function_call_output follow-up. Observe until the cumulative
provider-request count covers the follow-up and assert the recorded
function_call_output before finish; pin the contract in the workflow test.
* feat(e2e): record provider media content facts
* docs(mantis): list staged-media recipe in recipe index
* fix(mantis): isolate mock provider evidence
* fix(mantis): state provider facts as tamper-evident, not provenance-trusted
The mock sidecar makes provider request facts tamper-evident: candidate
code cannot alter or remove a recorded request. It cannot make them
provenance-authenticated — every process in the SUT container shares the
internal network, UID, and gateway config, so nothing distinguishes the
gateway flow from any other candidate-originated request. The proof
prompt and busy-queue recipe now state exactly that property instead of
calling the facts trusted.
* fix(mantis): redact unparseable provider request bodies
Media redaction walks the parsed JSON body, so a request that fails
JSON.parse fell back to logging the raw text — leaking base64 payloads
the redactor exists to strip. Unparseable bodies now log a bounded
byte-count marker instead; regression posts a malformed body carrying a
data URL and asserts the payload never reaches the record.
* fix(mantis): expose newest provider records through a seq-stamped tail
The lane's requests surface kept the first 100 provider records, so a
session longer than the window hid exactly the newest requests a proof
asserts on. The mock server now stamps each record with a producer-owned
absolute seq ordinal, and the lane reads a bounded 128-record tail —
mirroring the sibling botApiRequests window. Regression writes 130
records and asserts the tail keeps seq 3..130; it fails pre-fix.
Vincent Koc