chimera solve

The task to solve autonomously (omit with --approve/--deny).

Verification command (exit 0 == success).

Workspace root.

Override the model slug.

Max verify-or-revert attempts.

Max tool-calling steps per attempt.

Stop the whole run once this much has been spent (all attempts together).

Fraction of the model's window to spend on the prompt before compacting (e.g. 0.6).

Skip the planning step.

Skip Manager review.

Manager reviews via the cascade rubric.

Route deep-reasoning turns through fusion.

Tiered routing: weak -> gate -> mid -> gate -> fusion (cheap by default).

Gate tool calls through the governance kernel.

Per-session allowlist: only these tools (comma-separated).

Per-session denylist: drop these tools (comma-separated).

Track a capability ledger + review execution of tainted input.

Record trajectories for opt-in model evolution.

Don't auto-write a long-term memory fact on success.

Don't auto-propose a learned skill when a task recurs.

Run in an isolated git worktree; changes copied back only on success.

Give the agent an isolated Context Explorer for repo search (FastContext-style).

Give the agent spawn_subagent to delegate subtasks to isolated subagents.

Prepend a structural map of the workspace (files + top-level symbols) to the agent's context.

EXPERIMENT (study 20 B4): a cheap MODEL names the tool before each step and the executor is given only that tool. Reads a shallow context on purpose. Narrows only — an undecided router leaves the full list, and the run's receipt counts how often that happened.

EXPERIMENT (study 22 B4b): 'narrow' gives the executor only the routed tool (B4, measured worse); 'hint' keeps every tool and only suggests one for the step, with no ANSWER.

EXPERIMENT (study 24 M6): when the tool-loop breaker trips, hand the rest of the run to this stronger MODEL instead of stopping. A second trip stops as before. Off by default.

EXPERIMENT (study 24 M6 fork): with --escalate-on-tool-loop, copy the workspace to this DIR at the trip, before escalating — what stopping would have left. Off by default.

After a failed attempt, run a structured self-check that steers the retry (helps weak models).

Extract the task's atomic requirements and grade each attempt's coverage (catches dropped constraints).

With no --verify: generate executable pytest grounded in the task's requirements and use it as the gate (catches wrong code the coverage grade rubber-stamps). Measured on 78 labelled patches (bench/test_gate_two_sided): fails every wrong patch, and reverts 4 of 64 correct ones (6%) on a test of its own that is wrong — opt-in for that reason.

Model-role profile: economy | balanced | max. Puts a different model on each role (explore/plan/edit/review) drawn from the tier ladder. Routing is NOT yet shown to improve outcomes — see bench/role_routing/PREREGISTRATION.md.

Per-role model overrides: 'edit=vendor/slug,plan=vendor/other'. Roles: explore, plan, edit, review. Merges over --profile; a role left unset keeps --model. `verify` is not a role here — it runs a command and has no model to choose.

Comma-separated globs the file-writers may touch (e.g. 'src/**,*.py'). A write outside is refused — blocks an injected instruction from rewriting an unrelated file.

Log (arm, proxy=manager-judgment, reward=verified) per attempt to <home>/probe.jsonl for PROBE best-arm selection (see `chimera probe-select --from-log`). Needs --verify + a manager.

Reshape a long, rambling bug-report task into a salient-facts-first form (location/repro/expected-vs-actual/fix-hint) before planning. No-op on non-bug or short tasks.

Inject the stored ACE strategy playbook into context, then curate it from this run's outcome (closed loop).

Read learned skill cards back into context (the learn->use loop). Default follows settings.skill_cards; this overrides it per run.

With --fuse: sample K cheap answers per turn; escalate to fusion when they disagree (free confidence signal).

Model slug of a stronger, independent judge that grades hard-turn (retried) results before accepting them.

On a stall, rebuild the plan from accumulated failure causes (dual-ledger) instead of just nudging.

Show a failed attempt its own reverted diff, as a path not to retake.

On failure, leave the last attempt's edits on disk for an external grader (don't revert).

Fail an attempt that changed no file — for code tasks, an explanation is not a fix.

How a failed attempt's retry is briefed: generic (manager prose + verifier output) or targeted (a brief aimed at the classified failure).

Match repeated-failure signatures approximately, not byte-identically.

Machine-checkable success clauses, comma-separated: file_exists:PATH | file_contains:PATH:REGEX | answer_matches:REGEX.

Print live progress events (attempt/result/status) as the run proceeds.

Checkpoint this run under a thread id; re-run with the same id to resume after a crash.

Pause for human approval before finalizing a run that consumed untrusted content (needs --thread).

HITL accept: finalize a paused run as-is, by thread id (no task needed).

HITL ignore: discard a paused run by thread id (no task needed).

HITL respond: resume a paused run by thread id with --feedback guidance.

Guidance for --respond (fed back so the run tries again).

HITL edit: finalize a paused run with the corrected --answer, by thread id.

The human-corrected answer for --edit.