--- name: solve-efficiently description: Use for cross-file implementation, reproduced defects, bounded primary-source decisions, or authorized real-host workflows. Do not use for obvious one-file edits. --- # Solve Efficiently Choose the shortest path that proves the requested outcome. Read no other Cognitive Powers skill unless the first focused inspection leaves the boundary genuinely unclear. ## Locate plugin files Paths written as `scripts/` are relative to this skill's own directory. Paths written as `/...` are relative to the installed Cognitive Powers root: the directory that contains `skills/`, `scripts/`, and a `.codex-plugin/plugin.json` or `.claude-plugin/plugin.json` manifest. Resolve both from this skill's own location rather than guessing, and never copy plugin scripts into the target repository. ## Conservative agent control Obvious one-file work with a nearby check stays with the main agent; do not run an orchestration script merely to confirm that. After the first bounded discovery for non-trivial work, read the full skill and obtain the compact v2 input template with `/scripts/orchestration_policy.py --agent-plan-template 2 --json`. Fill that template and submit it with `--agent-plan --json`. Never read `orchestration_policy.py` to discover its contract. Follow its complete `solo`, `parallel-read-only`, `parallel-packets`, or `staged-verify` plan and keep its assignment IDs immutable. After each wave, evaluate the declared stop conditions without generating a replacement plan; if it becomes invalid, stop delegation and report degradation. Use schema v2 for durable work or quality claims. It binds assignments to the plan, requires an observed RED cycle before a separate test-writer, and requires an executable verification target before scheduling a verifier. Treat v1 worker results as compatibility data only; they cannot support a durable completion or quality claim. Use only the host's native agent tools. The main agent retains authorization, synthesis, integration, and final claims. Never depend on custom agent TOML discovery, a fixed model, a fixed swarm size, or hooks to create agents. A semantic planning-input error fails closed to `solo`; it never expands scope or grants write authority. Copy every enum from the template's `allowed_values` exactly. Do not invent aliases such as `write`, `workspace-write`, or `implement` for fields whose declared values are `change`. Do not add a `verifier` unit: provide the executable `verification_check` and let the planner create its fresh synthetic verifier. `selected_mode` is the planner decision. A new plan deliberately has `executed_mode=null`, `outcome=planned`, and `degradation=null`; only observed host execution may populate those receipt fields. If the main agent absorbs a delegated plan, report `executed_mode=solo`, `outcome=degraded`, and the exact degradation cause. Never present a selected mode as an executed one. ## Current-source decision When the checkout provides a decision brief, source policy, and primary-source files: 1. In one tool call, read the brief, policy, applicable `AGENTS.md`, and every supplied source. 2. Build an internal requirement matrix: each question or constraint maps to the exact claim needed, its source, version/date, and any unresolved fact. 3. Answer directly. Cite every decision-critical claim at claim level using the source identifier or primary URL required by the policy. 4. Cover every explicit requirement, including protocol headers, compatibility, defaults, retired behavior, and operational constraints. 5. Distinguish publication, rollout, deprecation, and retirement dates. Never convert one into another or invent a schedule the source does not state. 6. State uncertainty explicitly when the sources do not establish a fact. Use the brief's numbered order. Answer in the same number of compact numbered items, without a preamble or restating the brief, and stay under 350 words unless the brief explicitly requires more. Keep each claim and its citation in the same paragraph; do not separate them with a table, list, or fenced block. For requested protocol values or headers, state every exact name, value, requirement, and source marker together in one prose sentence; do not format protocol values as separate lines. Prefer the primary source's own terminology for status and behavior instead of replacing it with a looser synonym. Express obligations and lifecycle states with explicit labels such as `must`, `required`, `supported`, `unsupported`, or `retired`, not only indirect paraphrases. Before finalizing, silently check that every row in the requirement matrix appears once with its source marker. Do not browse when the supplied policy says the frozen primary sources are complete. Do not load another skill, run repository discovery, or make multiple source-reading passes. ## Bounded implementation When a contract or reproduction identifies a small source-and-test boundary: - Additive feature: discovery, patch, verification. - Existing defect: discovery, focused reproduction, patch, verification. The discovery call reads applicable instructions, the named contract or reproduction, matched implementation, and nearest tests together. Reproduce only an existing defect. Patch one coherent boundary. Verify the public behavior with the focused test and at most one distinct affected suite. Stop after it passes; do not add Git, compile, parser, diff, or unrelated test calls. ## Real-host workflow When `WORKFLOW.md` and `HOST_TOOL.md` are supplied, read them with applicable instructions in one call. Create the single declared public-action plan, execute it once, then run the fresh observation. Use only documented labels, roles, clicks, fills, selects, and keys. Do not call private APIs, edit persisted state, inject scripts, substitute a mock, or rerun a successful workflow. Report the exact action and observer receipt hashes. ## Escalation and handoff Read `../../skills/solve-efficiently/SKILL.md` only if focused inspection still leaves module boundaries, cause, or verification genuinely unclear. Otherwise report the result, exact checks already run, and any untested surface without re-reading files.