Add skill trigger pre-flight for prompts and A2A

Context

@tobiu surfaced a recurring swarm failure: Gemini-family sessions can forget to invoke applicable skills when the request uses ordinary action words instead of explicit skill names. Examples include test needing disambiguation between unit-test and whitebox-e2e, and review needing disambiguation toward pr-review when the target is a PR/diff/approval cycle.

This ticket follows the existing per-turn memory pattern in AGENTS.md §21: skill awareness belongs in turn-loaded memory when missing the trigger causes silent protocol failure.

The Problem

The current workflow skill table correctly documents lifecycle skill triggers, but it assumes the agent notices that a prompt maps to a skill. In practice, prompts and A2A messages often arrive as natural-language tasks:

• "test this"
• "review this PR"
• "re-review after my fix"
• "finalize/open the PR"
• "pick up this ticket"

When the agent misses that semantic mapping, the failure happens before the skill body is ever loaded. That is especially visible in A2A contexts, where a mailbox message may be treated as coordination metadata rather than as a prompt requiring the same trigger discipline as a direct user request.

This is not a lack of skill content. It is a trigger-reflex gap.

The Architectural Reality

• AGENTS.md:485 begins §21. Workflow Skills (when to invoke), the current per-turn awareness table.
• AGENTS.md:487 says skills must be invoked before the trigger action and that the table exists so the trigger -> skill loop fires reliably even when context pruning evicts skill files.
• AGENTS.md:502 explicitly documents the empirical failure mode: skipped skill invocation causes gate/protocol/template failures silently.
• AGENTS.md:504 currently says tactical/domain skills such as unit-test and whitebox-e2e remain discoverable via directory listing rather than per-turn awareness.
• AGENTS_STARTUP.md:61 already requires listing .agents/skills/ and reading relevant SKILL.md before multi-step testing/debugging/scaffolding, but this is boot/startup discipline, not a per-prompt or A2A-message pre-flight.
• Prior related issue #10365 established the compact workflow-skills table. This ticket narrows to the next missing layer: semantic trigger detection before prompt/A2A handling.

Five-stage challenge chain result:

1. Premise: Real and observed in current swarm coordination; the issue is skill under-triggering, not missing skill bodies.
2. Prescription: Add a compact AGENTS.md pre-flight. Do not create a new skill; the problem is detecting existing skill applicability.
3. Substrate: AGENTS.md per-turn memory owns this discipline because it must fire before planning and before opening any specific skill.
4. Consumer: Agent harnesses and A2A mailbox readers. The output must be terse, router-like, and survive context pruning.
5. Service Boundary: No MCP, Memory Core, or skill payload changes are required for v1.

The Fix

Add a compact "Skill Trigger Pre-Flight" near AGENTS.md §21, before the mailbox/turn-start sections. The rule should require every agent, before acting on a user prompt or A2A message, to identify the action and target object, then compare that pair against known skill triggers.

The pre-flight should explicitly cover ambiguous words by object:

• review + PR/diff/approval/LGTM/re-review => pr-review
• test + unit/spec/Playwright unit/file.spec => unit-test
• test + E2E/end-to-end/user flow/whitebox/Neural Link => whitebox-e2e
• assigned issue / ticket pickup => ticket-intake
• commit / push / open PR / finalize branch => pull-request
• regression / "used to work" / prior decision context => memory-mining

The wording should say "check prompt intent against skill triggers", not "check skill names", because users and agents usually do not name the skill explicitly.

Acceptance Criteria

• AGENTS.md §21 gains a compact Skill Trigger Pre-Flight applying to every user prompt and every A2A message.
• The pre-flight requires action + target-object disambiguation, not raw keyword matching.
• The pre-flight includes explicit examples for review, test, ticket pickup, PR finalization, and regression/history prompts.
• The text preserves Progressive Disclosure: it routes to existing SKILL.md files but does not inline heavy skill payloads.
• The text avoids implying that tactical skills should always be loaded; it only requires loading them when the semantic trigger matches.
• A short note clarifies that A2A mailbox messages count as prompts for skill-trigger purposes.
• git diff --check passes.

Out of Scope

• Creating a new skill.
• Editing .agents/skills/* payloads.
• Building an automated skill-trigger evaluation harness. Discussion #10320 covers broader skill trigger eval infrastructure.
• Changing MCP tool behavior or Memory Core mailbox semantics.
• Refactoring the full workflow skills table.

Avoided Traps / Gold Standards Rejected

• Naive keyword routing: Rejected. review can mean conceptual review, PR review, or design critique; test can mean A2A smoke test, unit test, or E2E test. The trigger must use action plus target object.
• New skill for skill-triggering: Rejected. The failure happens before skill selection; a new skill would itself need to be triggered and therefore would not solve the base problem.
• Inlining full tactical skill rules into AGENTS.md: Rejected. That breaks Progressive Disclosure and bloats turn memory. AGENTS.md should only contain the router/pre-flight.
• Treating A2A as lower priority: Rejected. A2A messages can assign reviews, request re-reviews, or hand off ticket work, so they must pass through the same skill trigger gate as user prompts.

Related

• #10365 — introduced the compact per-turn workflow skills table.
• #10376 — tightened A2A commentId hand-off reflexes for PR review / PR author workflows.
• Discussion #10320 — broader skill trigger evaluation infrastructure, deliberately out of scope for this small AGENTS.md router change.
• Current Memory Core session captured the analysis and recommendation.

Origin Session ID: ea6898ba-531f-4162-996b-99e7b235c120

Retrieval Hint: "skill trigger pre-flight prompts A2A review test unit-test whitebox-e2e pr-review AGENTS.md §21 Gemini under-triggering"