⛔ This skill MUST be read in full — not skimmed. Formal review gates depend on its workflow.
Skipping steps silently bypasses quality checks. Missing gates = undetected breakages.

Source: Canonical copy at skills/execution-intent/SKILL.md.

Execution Intent

Ontology: tortoise/docs/ONTOLOGY.md (v3.1, canonical) — fetch: gh api repos/daniel-ospina/tortoise/contents/docs/ONTOLOGY.md --jq .content | base64 -d (§5 = controlled vocabulary).

Overview

The execution intent flag is set once at brainstorming entry and read by every downstream skill in the pipeline. It enables three profiles for how skills handle gates, review cycles, fix policy, and sub-agent dispatch.

This skill is a reference — it does not run a procedure. Other skills cross-reference it to avoid duplicating intent semantics in each skill's prose.

Profiles

Profile	Philosophy	User pauses	Review cycles	Fix scope	Sub-agent dispatch
Fast (default)	Speed + quality	Normal interactive gates	Per-skill default (each skill's own bound — proportional, convergence-gated, or explicitly uncapped)	Per-skill default	Ask user
Autonomous	Measure twice, cut once	Zero	Full per-skill convergence, no early-exit cap	All P0/P1/P2 + verify each fix	Auto-dispatch by heuristic
Budget	Watch credits	Normal gates	Same as Fast	Same as Fast	Force in-chat (migrations excepted)

Default behavior: When no intent is set, fall back to Fast.

⚠️ These are profile BUDGETS, not caps. A profile selects how much of a skill's own review loop runs — it never imposes a ceiling, and never a bound tighter than the skill's own. The governing bound is always the skill's own — proportional-gates is canonical where the skill defines proportional tiers. The review loop is how quality is produced, so the default is to keep cycling until a clean exit, genuine convergence, or the skill's own stall/abort signal — stopping early because a count "feels high" is a bypass. (agent-infra#700: a hard-coded "4 cycles maximum" in AGENTS.md and the literal "4" in this table contradicted every convergence-gated skill and overrode the tighter bounds those skills set for themselves, causing verification to stop with fixes unverified.)

Reading the Intent

Skills read intent in this order:

1. Session context — if a prior skill in the chain (typically brainstorming) set EXECUTION_INTENT, use that value.
2. Worktree file — cat .pi/execution-intent 2>/dev/null returns one line of the form EXECUTION_INTENT=<value>. If file mtime is >24h, treat the file as absent.
3. Default to Fast — if neither source returns a recognized value (fast, autonomous, budget).

Recognized values are case-insensitive. Unknown values fall back to Fast.

Setting the Intent

Brainstorming asks the entry-point question and writes both:

• Session context — already-loaded for downstream skills in the chain (no I/O).
• .pi/execution-intent file — at the worktree root, for skills invoked outside the chain.

Pseudocode for setting:

mkdir -p .pi
echo "EXECUTION_INTENT=$intent" > .pi/execution-intent

The file is overwritten on each new brainstorming session. Intent is immutable per session, but a new session re-asks the question.

Lifecycle

• Set at brainstorming entry-point question.
• Cleared by commit-workflow on completion (success or abort): rm -f .pi/execution-intent.
• Auto-expires when file mtime is >24h old; stale files are treated as absent.

Tier × Intent Interaction

Intent modifies behavior within the tier the skill selects. Intent never forces a Micro-tier skill to run Complex-tier phases.

Worked examples:

• Micro + Autonomous: Codebase Explorer still skipped (Micro rule). Plan review is skipped at Micro (plan-review §When to Use); when plan review does run, the tier's bound in proportional-gates §Review Cycles is the governing budget — no profile-level ceiling. Fix P0+P1 (Micro skips P2 by default; intent does not override tier).
• Micro + Budget: codebase-read only — the micro proportional external-research trigger is skipped under Budget (issue #231 D3).
• Complex + Budget: code-review skips NVIDIA pattern scan (6 → 5 agents). writing-plans skips perplexity gate. All other Complex phases run normally.
• Standard + Budget: issue-scoping Phase 1.5 runs ≤ 2 external queries, codebase-first, fired only on P0-level gaps (new third-party dep / novel pattern with zero in-repo precedent) — the writing-plans Perplexity gate is the total session research budget (no double-charge). (issue #231 D3)
• Epic-tier + Budget: epic-plan research hooks and epic-scope granular queries defer to the epic research brief (codebase + brief only, zero external queries). (issue #231 D3)
• Complex + Autonomous: All Complex phases run; review cycles run to the skill's own bound (no early-exit cap — proportional tiers, convergence gates with a 10-cycle safety cap, or the skill's specific bound); sub-agents auto-dispatch.

Sub-agent Preamble

Every sub-agent dispatched in Autonomous or Budget mode receives this canonical block at the start of its prompt. Copy-paste the relevant variant; do not paraphrase.

## Execution Mode: Autonomous

- Fix ALL issues found (P0, P1, P2). Do not skip minor issues.
- After every fix: verify the fix worked. Run relevant tests.
- Confirm nothing regressed. If regression found, revert and try alternative.
- Log every autonomous decision with rationale, risk, and recovery path.
- Do not wait for user approval. Proceed autonomously.

## Execution Mode: Budget

- Favor cheapest tool options. Skip optional research calls.
- Use in-chat execution. Avoid sub-agent dispatch unless migration isolation required.
- Normal gates apply. Wait for user approval when required.

In Fast mode, no preamble is added — sub-agents operate as they always have.

Autonomous Decision Log Format

Decisions made under Autonomous are logged inline in the active design doc, plan comment, or PR description as:

**Autonomous decision:** Chose <X> over <Y> because <reason>.
**Risk:** <what could go wrong if this turns out to be wrong>
**Fix if wrong:** <recovery path>

The user reviews these on return and can override any decision by replying.

Disambiguation: brainstorming's "Path B"

Brainstorming's existing scope-routing "Path B: No taxonomy → AUTO-PROCEED" is a brainstorming-internal conditional (auto-proceed within that skill only). This is distinct from the session-wide Autonomous profile. The two coexist because the profile is set after the taxonomy gate.

Continue following the workflow as mandated by this skill. Do not skip steps.