Three changes, one of them a live bug.
**The bug.** `phase_summarizer` truncated agent output with `&s[..remaining]`,
a BYTE slice of arbitrary UTF-8. Agent turn output routinely carries arrows,
box-drawing and emoji, so a cut landing mid-character panics — taking down the
evaluation sweep for that phase, triggered by nothing more than an agent
writing a long enough line with a non-ASCII character at the wrong offset.
Replaced with `clamp_to_char_boundary`, tested across every cut offset of a
pure-4-byte string.
It is precisely the bug the clawhdf5 agents found and fixed in
`clawhdf5-migrate/src/validate.rs` this week — in our own code, in the path
that feeds the judge.
**Evidence budget** 60 KB -> 120 KB. Output headroom is worthless if the judge
cannot see the work: the verdict is only as good as what reaches it.
**Judge max_tokens** 2048 -> 16384. glm-5.3 is a reasoning model that spends
most of its budget on a `thinking` block before writing the verdict, and
running out mid-thought truncates it. A truncated verdict parses as empty and
FAILS CLOSED, burning one of the phase's passes on a judge that never answered
— how mission 01a00bbb lost one.
Measured ceiling: z.ai accepts max_tokens up to 131072 on both glm-5.1 and
glm-5.3 (131073 -> 400, "限制数值范围[1,131072]"), so 16384 is chosen for cost
and latency rather than capability, and only emitted tokens are billed.
Co-Authored-By: Claude Opus 5 <[email protected]>
Operator model policy: haiku ONLY for genuine yes/no questions; anything
requiring thinking is opus-5; coding is sonnet-5.
The mission AGENTS were running haiku, and nothing in the product said so.
`provider_alias_for` maps every `claude-*` binding onto the single alias
`claude_cli.default`, so a crew whose `model_binding` reads `claude-sonnet-5`
— as this deployment's does — still ran whatever that alias pointed at, which
was `model = "haiku"` in the runtime config. The binding is cosmetic; the
alias is the truth.
Measured consequence on mission 01a00bbb: the coding agents claimed six INT
items complete and had committed three, and the done_when judge caught it by
auditing git history against the claims.
Model assignments, by what the component actually does:
evaluator (done_when judge) haiku -> opus-5 reads evidence, audits it
against the repo, writes
guidance. The verdict is a
boolean; the work is not —
and this is the one component
whose failure mode is passing
work that was never done.
judge_model 4-8 -> opus-5
mission_refiner 4-8 -> opus-5 composition
phase_summarizer 4-8 -> opus-5 composition
swarm planner 4-8 -> opus-5 planning
subscription preflight head 4-8 -> opus-5
fallback chain head 4-6 -> sonnet-5 haiku stays BELOW it as a
last-resort link, never a peer
Every value stays env-overridable; only the shipped defaults move.
Co-Authored-By: Claude Opus 5 <[email protected]>
The research scenario passed 4/4 and the log underneath it said:
phase_summarizer: ... failed: anthropic 400 Bad Request: "Your credit
balance is too low to access the Anthropic API"
`phase_summarizer` and `mission_refiner` each built their own reqwest POST to
the Messages API with `x-api-key: $ANTHROPIC_API_KEY`. No audit of `.complete(`
call sites could have found them — they never touched a provider — so every
phase summary and every mission-brief refinement on this deployment had been
failing against an empty account while the phases themselves ran fine. The
summarizer even persisted an error row per phase, which is why nothing ever
retried loudly enough to notice.
Both now go through `subscription::complete_with_fallback`, so they inherit the
subscription-first credential choice, the 429 backoff, and the opus -> haiku ->
glm chain. The summarizer records the model that ANSWERED in
mission_phase_summaries.model rather than the one it asked for.
The guard is a source WALK, not a file list: any .rs under cm-api/src that
mentions the Messages API host or `x-api-key` fails the test. A hand-listed set
of files is exactly what let these two hide.
Co-Authored-By: Claude Opus 5 <[email protected]>
A phase used to complete when its topology_runs reached a terminal state --
purely structural. It marked itself done whether the agents produced the
artifact or wrote nothing at all, and it ran exactly once: execute_resumable's
skip(start) is resume, not repeat, and the only re-run path was a human
hitting the retry endpoint.
A phase can now carry `done_when`, a completion condition judged after each
pass against the evidence the agents actually surfaced. Not met and passes
remain -> the phase goes back to pending with iteration bumped, and the
verdict's reason is appended to the next pass's task text. That feedback is
what makes iteration converge rather than repeat -- the same mechanism /goal
uses, and that swarm.rs already uses for rejected work.
The evaluator runs on the SUBSCRIPTION model. CLAWMATES_EVALUATOR_MODEL
defaults to judge_model(), and a `runtime:<alias>` spec routes through
ZeroClawDriveExecutor -- a container agent on claude_cli, i.e. Claude Code on
the OAuth subscription, needing no platform API key. Same routing the door
governor uses.
Two deliberate departures from the governor's contract, both required:
- FAIL-CLOSED. Runtime::judge is fail-open and reads a verdict by
!contains("DENY"), so a model explaining why it *would* deny reads as
approval and an empty reply reads as approval. For completion that is
backwards: unsure must mean not done. The contract is swarm.rs's strict
JSON {"met","reason"} with .unwrap_or(false). Six tests cover the closed
paths -- prose, empty, missing field, non-boolean, transport error.
- judge_raw returns the raw reply; judge collapses to a bool too early to
carry a structured verdict.
Iteration scoping is the subtle part and has its own test: on pass 2 the
phase's own iteration is 1 but pass 1's completed run is still in the table,
so "are this phase's runs all finished?" must ask about the CURRENT pass or
that stale row closes out pass 2 the instant it is enqueued.
Evidence comes from phase_summarizer::collect_evidence, extracted from the
existing collect_material so the evaluator and the summary card cannot
disagree about what a phase produced.
done_when/max_iterations are promoted from phase config into columns (the
sweep filters on them every tick) and max_iterations is clamped to 20 at
insert -- the UI limits it too, but a runaway loop must not be one crafted
request away.
A phase with no condition completes exactly as before; that regression guard
is the first test in the file.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
New phase_summarizer background worker fires on any mission_phase
transition to a terminal state (completed/failed). Aggregates every
topology_runs.checkpoint.outputs[] + mission_tasks + mission_artifacts
bound to that phase and asks Claude Opus 4.8 to produce a structured
JSON card:
{ narrative, metrics, sources, tooling, next_actions }
Rendered inline on the mission page under each completed phase via
new PhaseSummaryCard component. Metrics grid is kind-specific:
research surfaces insights/sources/int_cards/artifacts, coding
surfaces cards_picked_up/commits/tests/issues, benchmark surfaces
regressions/improvements, security surfaces findings-by-severity.
New table: mission_phase_summaries (migration 0060), unique per
phase_id — regenerates on retry.
New endpoint: GET /api/missions/{id}/phases/{phase_id}/summary.
Model overridable via CLAWMATES_SUMMARIZER_MODEL. Reuses the
ANTHROPIC_API_KEY prod already carries for mission_refiner.
