agilab-intent-router

name	agilab-intent-router
description	Route terse AGILAB operator requests such as "do it", "review AGILAB", "next move", "update repos", "merge it", "check again", "release", and "cluster validation" into the right repo skills, safety mode, validation depth, and output contract using session-derived policy.
license	BSD-3-Clause (see repo LICENSE)
metadata	{"updated":"2026-07-07T00:00:00.000Z"}

AGILAB Intent Router

Use this skill first when a user request is short, ambiguous, or operational. The goal is to classify the request before acting, so terse commands map to the right AGILAB workflow instead of defaulting to a shallow answer or wrong repo operation.

Grammar

request      ::= intent target? depth? scope? output? mode? constraint?
intent       ::= review | implement | validate | release | sync | publish | explain | continue
review       ::= "review" | "audit" | "inspect" | "assess" | "analyze"
implement    ::= "fix" | "implement" | "add" | "address" | "improve" | "clean"
validate     ::= "check" | "test" | "validate" | "verify" | "prove"
release      ::= "release" | "ready for release" | "prepare release" | "publish"
sync         ::= "update repos" | "sync repos" | "merge" | "push"
publish      ::= "linkedin" | "youtube" | "teaser" | "thumbnail" | "article"
explain      ::= "why" | "what next" | "where" | "explain" | "inspire"
continue     ::= "do it" | "go on" | "next move" | "check again"
target       ::= "agilab" | "docs" | "skills" | "examples" | "release" | "cluster" | "ui" | "video" | "repo" | url
depth        ::= "quick" | "targeted" | "deep" | "full" | "detailed"
output       ::= "summary" | "review doc" | "patch" | "PR" | "commit" | "copy/paste"
mode         ::= "do not fix" | "fix it" | "push it" | "merge it" | "only report"
constraint   ::= "only agilab" | "no tests" | "from current repo" | "from another machine"

Precedence

1. Explicit user constraint wins over default routing.
2. Destructive, publish, merge, and push intents require current repo state inspection before action.
3. Review/audit defaults to report-first, not patch-first.
4. Fix/implement defaults to smallest patch plus targeted validation.
5. Release readiness requires authoritative workflow/tooling inspection.
6. Docs edits use canonical ../thales_agilab/docs/source then AGILAB mirror sync.
7. Cluster work must rediscover remote workers; never reuse a remembered IP.
8. If the previous turn established scope, do it, go on, and check again inherit that scope; otherwise inspect current repo state and ask only if action would be risky.
9. If one user message combines execution, validation, publish, merge, or follow-up planning requests, for example do it; validate; push if clean; merge it; then next move, treat it as a single ordered turn. Execute each explicit step only after its safety gate is satisfied, report the result, then give the next recommendation. This avoids extra prompt round trips while preserving current-state inspection before merge or recommendation.

Session-Derived Routes

Route these patterns before choosing tools:

User phrase family	Intent	Mode	Skills
review AGILAB, audit AGILAB, deep review, address this audit	deep audit	report-first; architecture-readiness gate before final audit; patch only if fix it follows	agilab-deep-audit, then agilab-testing if patching
do it, go on, fix it, next move	continue current objective	inherit previous scope	previous active skill plus plan-before-code for code changes
update repos, sync repos	safe repo sync	show command plan first; default AGILAB target set is agilab plus sibling thales_agilab when present	agilab-runbook
update skill, sync skills, make future agents do X	repo skill update	edit .claude, sync .codex, regenerate index	skill-creator, repo-skill-maintenance
ready for release, release it, badge, proof, PyPI	release verification	inspect release workflow/tooling first	agilab-release-verification
docs aligned, screenshot docs, link added, published docs	docs workflow	canonical docs then mirror	agilab-docs
cluster, remote worker, validate cluster	cluster validation	rediscover worker, split SSH/share/compute	agilab-testing, agilab-installer, agilab-runbook
teaser, youtube, thumbnail, linkedin	content packaging	no repo edit unless asked	agilab-product-reels for video assets
anything to inspire from <url>, think from <repo>, learn from <tool>	external benchmark	inspect source, extract transferable patterns, reject misfit, implement only if AGILAB-native	agilab-evidence-contracts, agilab-docs, plus domain skill

Behavior Contract

• For review AGILAB, always load agilab-deep-audit and define a review scope before reading code. If architecture context is not crystal clear, load ARCHITECTURE_FOUNDATIONS.md and run the deep-audit preflight before writing final findings.
• For fix it after an audit, convert prioritized findings into a patch plan; do not silently patch unrelated findings.
• For update repos, never touch repos outside the current allowlist unless the user names them. In AGILAB maintenance, that allowlist includes the active agilab checkout and sibling thales_agilab when present. Show concrete git -C commands for both before execution, or explicitly report why one is missing, dirty, unsafe, or out of scope.
• For merge <branch> when the branch may belong to a sibling checkout, verify the actual repository, PR state, remote branch, and commit ancestry before acting. If the PR is already merged and the branch tip is an ancestor of the target remote main, report the merge as complete even if a stale local branch remains checked out. Treat deleting that local branch, switching branches, or cleaning a dirty checkout as a separate cleanup operation that requires an explicit non-destructive plan or explicit discard approval.
• For all clean? and follow-up cleanup requests such as clean it, report remote alignment, working-tree status, stashes, registered worktrees, and local branch refs separately before deleting anything. If cleanup is requested, preserve recoverability first: archive dirty worktree diffs, untracked files, stash patches, and local non-main branch refs in an out-of-repo cleanup archive or bundle, then remove stale worktrees, gone branches, and temporary stashes only after the archive exists. Finish by rerunning the same final-state stack so the close-out distinguishes the active checkout from the wider workspace.
• For do it after a proposal, execute the proposed action if repo state is safe. If state changed unexpectedly, report the changed branch/files first.
• For combined commands such as do it; validate; push if clean; merge it; then suggest next move, complete the ordered chain in one response when safe. A merge step still requires the current PR/branch to be clean, synchronized, and passing required checks; stop and report the blocker instead of asking for a second merge it or next move turn.
• For terminal checks in terse execution chains, prefer tokki run -- <command> when Tokki can preserve command behavior. Use raw commands for interactive flows, exact structured stdout, live streaming logs, or intentionally unwrapped behavior.
• For check again, verify the authoritative source for the prior claim: current repo state, workflow status, docs mirror, PyPI/GitHub release, or cluster discovery as appropriate.
• For next move, inspect current repo state and relevant workflow/tooling; do not answer only from memory.
• For external product URLs, browse or inspect the current source first, then separate reusable pattern from product-specific scope. If the next user says do it, implement the strongest AGILAB-native primitive with schema, tests, docs, and validation instead of adding vague roadmap copy.

Regenerating The Policy From Sessions

Use the synthesis tool to mine local sessions and rollout summaries into a redacted local artifact:

uv --preview-features extra-build-dependencies run python tools/session_intent_synthesis.py \
  --sessions-root "$HOME/.codex/sessions" \
  --memory-root "$HOME/.codex/memories/rollout_summaries" \
  --output test-results/session_intents.json

The output is evidence for updating this skill. Do not commit raw session transcripts or unreviewed local synthesis artifacts.