Cobalt Cockpit v0.3 implementation report

Starting SHA: 2008993 (Release Cobalt Cockpit 0.1.1). Plugin and marketplace version are now 0.3.0. The v0.2 and v0.3 changes remain in the working tree; the release commit is described at the end but was not pushed.

Goal#

Make reasoning effort dynamic and task-aware instead of fixed per tier, across the existing Opus 5.5 / Sonnet 5.5 / Haiku 5.5 architecture, without introducing Fable, JEV, OpenRouter, external inference routing or any new orchestration dependency. Preserve the v0.2 Elastic Swarm architecture.

Three ideas kept apart#

REQUESTED   what the commander (or AUTO) asked for: a level, or AUTO
SELECTED    for AUTO, what the task facts imply; for a manual request, the level itself
APPLIED     what the model's capability actually allows, after a fallback

A requested level the model cannot honour is not pretended. The applied level is the closest supported one, the fallback is named (EFFORT FALLBACK), and requested and applied effort are recorded separately in the Run Ledger.

Policy (hooks/effort.ts)#

A pure module: arithmetic and naming over facts the commander, the task and the engine already reported. It never calls a model, never touches $ and never sets a global setting. The one place effort is applied is turn.step, which rewrites a request's effort for a single turn, so two agents side by side can ask for different levels.

• AUTO selection: extractive (inventory, extraction, classification, summary) → low; implementation/simpleTest → low; normal engineering (feature, bugfix, refactor, dependencyMap, inspection, triage, coordination) → medium; architecture, verification, investigation, recon, ambiguity/uncertainty/coupling → high; debug, security, concurrency, migration, integrationFailure, high-risk architectural work → xhigh; extraordinary → max. A task that names no class uses its tier baseline (Haiku low, Sonnet medium, Opus high). A named class is capped at its tier's own ceiling.
• Capability: configured override > runtime observation > declared baseline. Declared: Opus …xhigh,max, Sonnet …xhigh, Haiku …high, unknown family the conservative common core low|medium|high. Nothing assumes max without an observation.
• Observation: the engine reports the level it really applied. Honoured (actual ≥ requested) adds the level; a downgrade removes it. Nothing is invented.
• Resolution: chooseEffort applies the operator ceiling first (so the reason names the ceiling rather than a capability), then clamps to capability, preserving requested exactly.
• Escalation: escalationFor moves one step — effort within the tier first, then Haiku → Sonnet → Opus — with a bounded budget and no jump to the top.

Host wiring (hooks/register.tsx, hooks/ledger.ts, hooks/swarm.ts, hooks/orchestra.ts)#

• turn.step builds facts from the task's role and mode, takes the commander's named level if present, otherwise AUTO (or the fixed level in MANUAL), and rewrites model and effort on the request. It records resolution.selected as the requested effort and the clamped level as applied, with routingReason and fallbackReason, and folds the applied level onto the swarm task via setAppliedEffort.
• classic.PostToolUse is the only place the engine's real applied level surfaces. It warns when the applied level differs from the recorded requested level and folds the fact into the per-model observation map. The comparison reads requestedEffort; this was corrected during implementation because classicTelemetry runs first and overwrites effort with the engine's level, which would otherwise make both the warning and the observation inert.
• The meter, the wide HUD (model · effort), the pane REASONING row, the swarm EFFORT STEPS row, the Run Ledger (requested/applied/observed) and /cockpit version (mode, ceiling, observed capability) all expose the policy.
• The composed system prompt appends effortPolicyText, and the swarm tool gained effort / effort_reason on assign and escalate.

Configuration#

New userConfig fields, both strings with declared options: reasoningMode (AUTO default, MANUAL) and maxEffort (max default, down to low). Operator-declared per-model capability (modelEffort) is parsed by the module but has no manifest field, because userConfig values are scalars or string lists; it remains an internal override.

Tests and verification#

New tests/effort.test.ts covers the pure policy (vocabulary, declared capability and observation, AUTO selection, ceiling and fallback, escalation). New tests/effort-host.test.ts covers the real hooks: AUTO levels reaching the engine, a commander-named level, the operator ceiling, MANUAL, capability fallback with the EFFORT FALLBACK warning and the observed map, /cockpit version, the HUD effort label, the pane REASONING row and the composed prompt.

Live acceptance#

The plugin was loaded from source with claude --plugin-dir <repo> (CLI 2.1.293) for two bounded, non-interactive turns:

1. Public defaults (orchestration off), --model opus: a real turn completed on claude-opus-5-5, exit 0, no stderr. This verifies the plugin loads and does not break a live turn.
2. Orchestration and dynamic reasoning enabled through a temporary --settings pluginConfigs file: the model confirmed its system context contained Reasoning mode is AUTO and quoted the exact sentence The ceiling for any request is MAX; a higher request falls back and is recorded., exit 0. This verifies the userConfig fields parse, prompt.compose injects the policy and the orchestration/effort path is active in the real host.

Installation state#

The v0.3.0 tree was synced over the user-scope directory marketplace install (a 0.2.0 backup was taken first), its local marketplace identity cobalt preserved, and claude plugin update cobalt-cockpit@cobalt moved the installed plugin from 0.2.0 to 0.3.0. claude plugin list reports 0.3.0 read from that folder; claude plugin details reports the six agents (explorer, researcher, reviewer, scout, utility, worker).

Known limitations#

• Live per-tier subagent effort (Haiku low/high, Sonnet medium/high) was not exercised as live model runs. Subagent model/effort rewriting is verified through the real hook boundary in tests/orchestration.test.ts, tests/effort-host.test.ts and tests/public-release.test.ts with synthetic engine responses; the live acceptance exercises the main loop and the policy injection.
• Capability is learned only from what the engine reports it applied. A transient downgrade is treated as an unsupported level and removes it from the observed map, as documented.
• modelEffort has no manifest field; it can only be set programmatically or in tests.
• The v0.2 limitations on host stop APIs, wildcard serialization, macOS rendering and AUTO backpressure are unchanged.

Files changed in v0.3 (relative to the v0.2 tree)#

• .claude-plugin/plugin.json, .claude-plugin/marketplace.json — version 0.3.0, reasoningMode / maxEffort options, description.
• CHANGELOG.md, README.md — v0.3 notes, dynamic-reasoning section, configuration rows.
• hooks/effort.ts (new), hooks/register.tsx, hooks/ledger.ts, hooks/swarm.ts, hooks/orchestra.ts, types/index.d.ts.
• tests/effort.test.ts (new), tests/effort-host.test.ts (new), tests/world.ts.
• scripts/audit-public.py — expected version 0.3.0.
• docs/implementation-v0.3.md (this file).