Execution Log: the cursor adapter passes the model as -m, an option cursor-agent does not have

Append-only log of progress for the user's visibility.

Phase transitions

Phase	Entered	Reviewed/approved by	Notes
phase-selection	2026-09-14	—	Risk tier 2 (autonomous-complete): one string constant in an adapter, a rule written into a capability doc, and tests. No sensitive path is touched — nothing under .the-loop/**, no *schema*, no .github/workflows/**, no auth/secret/credential path — and no trust boundary moves (evidence/security-review.md). Tier 2 is below the tier-4 named-security-sign-off threshold. brainstorming skipped — the reporter measured the CLI and named the fix; design-critic-review not selected (critics: [] in this repository, so a round would be unavailable). No authorized the-loop execute reaches this cloud session, so the selection is recorded here and every phase is walked by hand
requirements-definition	2026-09-14	pending — this branch's PR	bugfix.md — two requirements, three abuse cases. Root cause confirmed by building the argv, not by reading the ticket
design	2026-09-14	pending — this branch's PR	design.md — one constant, four decisions, decision-125
test-planning	2026-09-14	pending — this branch's PR	testing-plan.md — thirteen rows, five applicable; every n/a carries its reason
tasks-breakdown	2026-09-14		tasks.md — six tasks, red root first
implementation	2026-09-14		On claude/github-issue-360-uav3nz. TDD: evidence/red.md captured before the fix
verification	2026-09-14		evidence/verification.md — every applicable row plus make check; evidence/security-review.md — three abuse cases, no findings
needs-review	2026-09-14		PR raised with the R10 briefing. Tier 2 permits autonomous completion after the review loop; the merge is still the owner's
complete

Pull requests

PR	Scope / tasks	Status
#361	tasks 1–6: the whole work item	open

Progress entries

2026-09-14 — the ticket's one-line fix, and the second surface it also breaks

• Phase: requirements-definition → design
• Did: confirmed the root cause by building the argv rather than by reading the adapter (oneshot_argv('review this', model=…) → […, '-m', 'gpt-5.6-sol']), then traced every reader of model_flag.
• Found, and it widened the spec without widening the diff: the critic round is not the only victim. the-loop models check builds the same one-shot argv (modelprobe.py:158) and reads a non-zero exit as refused — the harness said no. cursor-agent exits non-zero on -m before it has an opinion about any name, so every declared model comes back refused on cursor, is cached, and is then withheld from the phase-selection checklist: a human cannot pick a cursor model at all. A parse error about the-loop's own flag was being recorded as the vendor refusing the operator's model.
• Also checked, and it is not affected: interactive dispatch. cursor-agent has no pre-assignable session id, so the adapter's interactive_* methods raise UnsupportedRunnerError and no cursor session is ever spawned. The blast radius is critics and the probe, and nothing else.
• Why the repository did not catch it: the two assertions that named the flag compared the adapter's own constant to itself (("-m", …) on both sides), and the one test that ran the whole critic argv had -m written into it. The suite agreed with the bug. The new test asserts the argv from oneshot_argv, which is what reaches the process.
• Next: the red root, then the constant.

2026-09-14 — red, one line, and a cache that heals itself

• Phase: tasks-breakdown → implementation → verification
• Did: tasks 1–6. Wrote the two new assertions and captured the failing run (evidence/red.md); changed model_flag to --model with the reason beside it; corrected the three assertions that had pinned -m; added the verdict-cache test; wrote the rule into docs/capabilities/review-loop.md and decision-125.
• The migration that isn't: every cursor model probed during the -m era is cached as refused, and a standing refused withholds a choice. Nothing has to clear it — a verdict records the digest of the argv it was taken against, and VerdictCache.get drops one whose digest no longer matches (modelprobe.py:225). Changing the flag changes the digest. That is asserted (T3) rather than assumed, in both directions: the cached refusal still withholds under the old digest, and no longer does under the new one.
• Checkpoint/tests: red first, then green. make check: 3605 passed, 1 skipped (3603 before — the two added tests); ruff, ruff format, pyright, validate_config clean; markdownlint 0 errors over 1103 files.
• Self-review, three passes. Pass one found the new verdict-cache test asserting only that the choice is offerable after the fix — which a broken offerable would also satisfy — so it now asserts the converse first: still withheld under the old argv's digest. Pass two re-read the capability bullet, which ended "claude is --model; cursor-agent is --model" — two clauses saying one thing; now "Both shipped adapters are --model." Pass three checked the claim in design.md that the duplicated-flag warning starts working: it does, but it is config_findings over harnesses[<name>].args (session dispatch), not over the critics[].args the ticket's workaround uses. Corrected rather than dropped, because the distinction is the reason the workaround never warned.
• Security review: the built-in skill, scoped to this branch's own changes — its origin/HEAD diff resolved to the whole previous release on this fresh clone, which is another work item's paper trail. No findings; the three abuse cases are all "unchanged boundary", each mapped to an existing negative test that still passes (evidence/security-review.md).
• Out of scope, noted for the owner: two pieces of drift on main, reverted here rather than carried into this diff. uv.lock still records the-loopy-one 15.0.0 while cli/pyproject.toml is at 16.0.0 after the release bump, so any uv run rewrites one line of the lockfile and every contributor gets a dirty tree; and cli/tests/test_graph_drive_integration.py writes a portable record into the checkout's own .the-loop/ (.the-loop/portable/github-octo-repo-15.json changes whenever the suite runs) — the residue issue-339's log already flagged. Neither belongs in a bugfix about a flag.
• Next: the PR and its briefing.

Verification results

Only when this work item declared test-planning away (issue-179). This item kept the plan, so its results live in testing-plan.md §Verification results and this section stays as the template left it.

Design critic review

Not selected at phase-selection — this repository declares no critics (critics: []), so a round would be recorded unavailable and count for nothing. The section stays as the template left it.

Review cycles

Cycle	Type (self/critic/security)	Reviewer	Outcome	Link
1	self	the-loop	new findings — the one-sided cache assertion, the duplicated capability clause, the harnesses[].args vs critics[].args mix-up; all three fixed	this log, entry 2
2	self	the-loop	zero (converged)	this log, entry 2
3	self	the-loop	zero (converged)	this log, entry 2
4	critic	—	unavailable — no critic is configured in this repository (critics: []); does not count toward reviews.criticReviewCount	.the-loop/cli-config.yaml
5	security	built-in security-review skill	zero — no findings	evidence/security-review.md

Security review (gate)

• Mechanism: the harness's built-in security-review skill, scoped to this branch's own changes (see evidence/security-review.md for why the tool's own diff range was wider).
• Outcome: pass — no findings. The three abuse cases of bugfix.md §Security considerations are all "unchanged boundary", each mapped to an existing negative test that still passes. The change restores a review control that was silently not running rather than relaxing one.
• Human sign-off: n/a — risk tier 2, below the tier-4 threshold.

Final validation evidence

Acceptance criterion	Proof
R1.1 — a resolved model is passed as --model <name>	test_cursor_oneshot_argv_uses_the_long_model_flag, test_a_model_resolves_through_the_adapters_flag (T1)
R1.2 — the full critic argv for harness: cursor + model:	test_builtin_harness_derives_argv_from_the_adapter (T2)
R1.3 — no model, no flag	test_oneshot_argv_without_a_model_is_the_plain_one_shot_run, untouched and passing (T1)
R1.4 — fails before, passes after	evidence/red.md → evidence/verification.md
R2.1 — the model_flag rule in the capability doc	docs/capabilities/review-loop.md §Current behaviour
R2.2 — a History row tracing it to issue-360	docs/capabilities/review-loop.md §History, first row

make check green on the final tree: 3605 passed, 1 skipped; 0 type errors; 0 markdown errors. Full transcript in evidence/verification.md.

Capability docs

Capability doc	What changed	History row
review-loop.md	§Current behaviour gains the rule that an adapter's model_flag is a spelling the harness's own --help lists — the long form where a CLI offers both — with the reason (the flag reaches a process the-loop does not parse for, so a wrong one is a dead run, not a degraded one) and the current value for both shipped adapters	issue-360, linking this spec and decision-125

No other capability doc changed. cli.md describes the-loop models check as the verb that probes a declared choice, which is still exactly what it does — the fix changes which flag the probe sends, not what the command promises.

Documentation

Document	What changed
decision-125	New. Where an adapter's flag spellings come from, why they are not derived from the binary at runtime, and the four alternatives rejected — including editing the locked issue-358 spec that records the wrong belief
decisions.md	The index row for decision-125

No user-facing page under docs/cli/ or docs/config/ changed, and the reason is that none of them ever named a per-harness flag: docs/config/cli/critics-options.md says the model is "passed via the harness's own model flag" and docs/config/cli/harnesses-options.md says the name is passed to it "verbatim". Both statements were true before this change and are true after it — the defect was that the adapter's value did not match the CLI, which is a code fact, not a documented one.