triage

name	triage
description	Routes a natural-language prompt to the right workflow skill. Reads the prompt, inspects workflow state, classifies intent against the full skill catalog, proposes the best-fit skill with rationale, then tells the user which command to type. Use for: /triage, 'what should I run', 'which skill fits this', 'route this', 'pick the right command for me', 'I'm not sure what to do next'. Never invokes skills directly — always propose-only.
model	haiku

Triage — Prompt-to-Skill Router

Portable intent doc: quoin/core/skills/triage.md

§0 Model dispatch (FIRST STEP — execute before anything else)

This skill is declared model: haiku. If the executing agent is running on a model strictly more expensive than the declared tier, you MUST self-dispatch before doing the skill's actual work.

Detection:

• Read your current model from the system context ("powered by the model named X").
• Tier order: haiku < sonnet < opus.
• Sentinel parsing: the user's prompt is checked for the [no-redispatch] family.
  • Bare [no-redispatch] (parent-emit form AND user manual override): skip dispatch, proceed to §1 at the current tier.
  • Counter form [no-redispatch:N] where N is a positive integer ≥ 2: ABORT (see "Abort rule" below).
  • Counter form [no-redispatch:1] is reserved and treated as bare [no-redispatch] for forward-compatibility; do not emit it.
• If current_tier > declared_tier AND prompt does NOT start with any [no-redispatch] form: Dispatch reason: cost-guardrail handoff. dispatched-tier: haiku.

Pre-dispatch 1M check (IVG-90 Layer 1+2):

• Run: python3 QUOIN_HOME/scripts/dispatch_config.py --decide --tier <declared_tier> --verbose where <declared_tier> is the tier declared for this skill (e.g. "sonnet" or "haiku", as shown in the dispatched-tier line immediately above).
• If the command returns "safe-path" on line 1: Read the reason token from line 2 (config|cache|probe). Emit the one-line advisory (verbatim, substituting with the line-2 token): [quoin: 1M-unsafe declared-tier per <reason>; running SAFE PATH without dispatch] Then proceed to §1/§0c at the current tier (treat as if [no-redispatch] were present). Do NOT call the Agent dispatch. Do NOT call AskUserQuestion.
• If the command returns "dispatch" on line 1, OR if the script is missing / errors: Continue to the Agent dispatch call below (today's path — fail-OPEN).

  Spawn an Agent subagent with the following arguments:
    model: "haiku"
    description: "triage dispatched at haiku tier"
    prompt: "[no-redispatch]\n<original user input verbatim>"
  Wait for the subagent.

  Cache the safe result (best-effort):
    python3 __QUOIN_HOME__/scripts/dispatch_config.py --write-cache --tier <declared_tier> --result safe
  (Fail-OPEN: if the script errors or is missing, silently skip and continue.)

  Return its output as your final response. STOP.
  (Return the subagent's output as your final response.)

Abort rule (recursion guard):

• If the prompt starts with [no-redispatch:N] AND N ≥ 2: ABORT before any tool calls.
• Print the one-line error: Quoin self-dispatch hard-cap reached at N=<N> in triage. This indicates a recursion bug; aborting before any tool calls. Re-invoke with [no-redispatch] (bare) to override.
• Then stop. Do NOT proceed to §1.

Manual kill switch:

• The user can prefix any user-typed slash invocation with bare [no-redispatch] to skip dispatch entirely (e.g., [no-redispatch] /triage).
• This is the user-facing escape hatch and intentionally shares syntax with the parent-emit form: a child cannot tell whether the bare sentinel came from the parent or the user, and that is by design — both paths want the same proceed-to-§1 outcome.
• Use this only when intentionally overriding the cost guardrail (e.g., for one-off debugging on a different tier).

Fail-graceful path with error-class triage (per architecture I-01):

• If the Agent tool returns an error during dispatch, classify the error message text BEFORE proceeding:

• Error classification:

  • Worktree-class: the error text contains the substring Cannot create agent worktree, OR (the substring worktree AND the substring not in a git repository). This is recoverable — the harness tried to create a git worktree for isolation and the project root is not a git repo. Continue to Worktree-class branch.
  • Other-class: any other tool error, exception, or harness rejection — skip to Other-class path below (existing fail-OPEN behavior).

• 1M-credit-class: if the error text contains the substring Usage credits required for 1M context: This is the 1M-context credit mismatch (IVG-89). The parent session carries the context-1m-2025-08-07 beta header which propagates to all subagent calls; the declared-tier model lacks 1M credits. Detection via model-name is impossible; this post-dispatch error string is the only reliable signal. Emit (verbatim): [quoin: 1M-context credit mismatch on <tier> subagent dispatch; proceeding in-session at parent tier — run /model to switch this session to standard context for a permanent fix]

  Cache the unsafe result (best-effort):
    python3 __QUOIN_HOME__/scripts/dispatch_config.py --write-cache --tier <declared_tier> --result unsafe
  (Fail-OPEN: if the script errors or is missing, silently skip and continue.)

  Then proceed to §1 at the current tier (treat as if `[no-redispatch]` were present).
  Do NOT retry the Agent dispatch. Do NOT call AskUserQuestion.

• Worktree-class branch: Worktree creation is hook-driven and cannot be skipped by omitting a parameter. Use the AskUserQuestion tool to present the user with one option: (c) proceed-current-tier — Skip dispatch, proceed at the current (more expensive) tier. This is the only available recovery path. Question header: Subagent dispatch failed (worktree creation). Proceeding at current tier. Note for the user: "Worktree dispatch failed and no retry mechanism is available — worktree creation is unconditional in this harness. Proceeding at current tier."

• Other-class path (also: worktree-class after user acknowledges c): Do NOT abort the user's invocation. Emit the bare warning (verbatim): [quoin-stage-1: subagent dispatch unavailable; proceeding at current tier] If this path was reached via a worktree-class error, ALSO emit the classification line (second, separate): [quoin-stage-1: error-class=worktree; user-choice=c; proceeding at current tier] Then proceed to §1 at the current tier (fail-OPEN per I-01).

Otherwise (already at or below declared tier, OR prompt has [no-redispatch] sentinel, OR dispatch unavailable): proceed to §1 (skill body).

Summary

/triage is a lightweight routing skill. It reads a user's natural-language prompt, inspects the current workflow state on disk, scores each candidate skill against an embedded catalog, and presents a ranked proposal with rationale. After confirming the best-fit skill, /triage always tells the user which command to type — it never invokes skills itself. This applies to every skill, not just /implement and /end_of_task. The propose-only posture for all 19 routable skills is intentional and permanent: no Skill tool calls happen from /triage at any point.

---

Session Bootstrap

/triage is a routing skill. It does minimal setup:

1. CLAUDE.md is injected by the harness — no explicit read is needed.
2. Cost recording — explicit-only: Record a cost-ledger entry ONLY if the user's prompt explicitly names a task folder (e.g., /triage task=auth-refactor or the prompt contains the exact folder name). In all other cases — including when exactly one non-reserved task folder exists — skip cost recording silently. This conservative policy prevents false ledger writes in multi-task-folder contexts.
3. Session state: /triage does NOT write .workflow_artifacts/memory/sessions/ files. Routing is not "meaningful work" in the session-state rule sense.
4. Trigger-phrase stripping (see Step 1 below) happens before any scoring.

---

Process

Step 1: Strip trigger phrases

Before scoring, remove any exact match of the /triage description's trigger phrases (lowercased, trimmed) from the prompt. The trigger list to strip:

• "what should i run"
• "which skill fits this"
• "route this"
• "pick the right command for me"
• "i'm not sure what to do next"
• "/triage"

Stripping prevents the invocation phrase from polluting Signal B scoring. After stripping, normalize whitespace. The resulting text is the scoring prompt.

Step 2: Collect workflow-state signals

Run the following checks. All are best-effort and read-only — if a command fails or a file is missing, treat the signal as "unknown" (0 points). Never error out. Evaluate once; do not re-evaluate mid-conversation. Active task folder checks always exclude finalized/, memory/, and cache/ subdirectories.

Signal	Detection	Skills boosted	Skills suppressed
No .workflow_artifacts/ at project root	ls .workflow_artifacts/ 2>/dev/null returns empty/error	/init_workflow +2	All task-execution skills −1 each
repos-inventory.md missing or >7 days old	ls .workflow_artifacts/memory/repos-inventory.md 2>/dev/null absent, or mtime >7 days (stat -f %m macOS / stat -c %Y Linux, fallback: treat absent as stale)	/discover +2	/architect, /plan, /thorough_plan −1 each
At least one active task folder exists	ls -d .workflow_artifacts/*/ 2>/dev/null filtered: exclude finalized/, memory/, cache/	/review, /implement, /gate +1 each	/init_workflow −2
Plan found at any depth under active task folder(s)	find .workflow_artifacts -maxdepth 4 -name current-plan.md -not -path '*/finalized/*' -not -path '*/cache/*' -not -path '*/memory/*' 2>/dev/null returns any result	/implement +2, /critic +1 (if no critic-response-*.md found near it)	/plan −1, /thorough_plan −1
Plan found AND recent commits AND no review-*.md	Plan found (above) + git log --since=1.day --oneline 2>/dev/null non-empty + find .workflow_artifacts -maxdepth 4 -name 'review-*.md' -not -path '*/finalized/*' 2>/dev/null returns nothing	/review +2	/plan −1, /thorough_plan −1
review-*.md with verdict APPROVED	grep -rl "APPROVED" .workflow_artifacts/*/review-*.md 2>/dev/null non-empty	/end_of_task +2	/review −1
Uncommitted changes on current repo	git status --porcelain 2>/dev/null non-empty	/implement +1	/end_of_task −1 (unless APPROVED)
On main/master with clean working tree	git status 2>/dev/null clean + git branch --show-current 2>/dev/null is main or master	/start_of_day +1, /thorough_plan +1	/end_of_task −2
Today's daily cache exists	ls .workflow_artifacts/memory/daily/$(date +%Y-%m-%d).md 2>/dev/null found	/end_of_day +1	/start_of_day −1
No daily cache for today AND no session file for today	Daily cache absent + ls .workflow_artifacts/memory/sessions/$(date +%Y-%m-%d)-*.md 2>/dev/null absent	/start_of_day +2	—
Active session file with Status: in_progress	grep -l "Status: in_progress" .workflow_artifacts/memory/sessions/$(date +%Y-%m-%d)-*.md 2>/dev/null non-empty	/implement +1, /review +1	/init_workflow −2
It's Friday (local date)	date +%u 2>/dev/null == 5 (uses local timezone — correct; date +%u respects the system clock)	/weekly_review +1	—
Knowledge cache stale (>7 days)	stat -f %m .workflow_artifacts/cache/_staleness.md 2>/dev/null (macOS) or stat -c %Y (Linux) > 7 days, or file absent	/discover +1	—

Step 3: Classify against the skill catalog

Apply the three-signal scoring algorithm:

Signal A — Explicit command. If the scoring prompt contains a literal /<skill-name> token matching any of the 19 routable skills (case-insensitive), that skill gets score 3 and skip directly to Step 4 proposal. No further scoring needed.

• Special rule: if /revise-fast is matched via Signal A, substitute /revise in the proposal with note: "/revise-fast is an internal variant; proposing /revise instead."
• /revise-fast is never included in any ranked list or ambiguous candidate display.

Signal B — Trigger phrase match. Match the scoring prompt (lowercased, whitespace-normalized) against the Trigger phrases column in the catalog (Section 6). Each exact phrase match = 2 points. Each partial primary keyword match = 1 point. Sum per skill.

Signal C — Workflow state. Apply the boost/suppress values from Step 2 to each skill's running total.

Step 4: Rank candidates

Sort skills by total score (Signal B + Signal C) descending.

• Single winner: Top skill score ≥ runner-up by ≥ 2 points → single-candidate flow.
• Ambiguous: Top N candidates within 1 point, N ∈ {2, 3} → ambiguous-candidate flow.
• Decline: No skill scores above 1, OR N > 3 candidates tied, OR top candidates all have "High" impact with no disambiguating state signal → decline flow.

§V Reconcile (read-only — no side-effect check)

Before surfacing any task/PR status, run python3 __QUOIN_HOME__/scripts/verify_claims.py --reconcile-tasks --project-root <project-root> and derive the displayed status from the reconcile table, not from a cached narrative alone. If the reconcile exits 8, surface the contradiction rather than silently reporting the narrative version.

Step 5: Present proposal

See Section 9 (Ambiguity Handling) for the exact dialogue format.

Step 6: Handle confirmation

Once the user confirms any skill from any flow:

1. Print: → Type /<skill-name> to proceed.
2. Stop. Do not call the Skill tool. This applies universally — no exceptions.

For decline/negation replies, apply the one-question cap (Section 9).

---

Section 6: Skill Catalog

All 19 user-facing routable skills. /revise-fast is excluded (internal, not user-facing).

Skill	Primary keywords	Trigger phrases	State signals that boost (+)	State signals that suppress (−)	Notes
/architect	architect, system design, design, explore	"design the system", "architecture", "how should we build this", "explore the codebase", "technical exploration", "cross-repo"	architecture.md missing + task folder exists (+1)	Plan already exists (−1)	Large tasks; use before /thorough_plan
/capture_insight	note, insight, remember, log, save, pattern, gotcha	"note this", "remember this", "log this", "save this as a lesson", "remember that", "note that", "I want to capture"	Active session in_progress (+1)	—	Low-impact, quick logging
/continue_work	continue, resume, revive, prior session, last session, where was I	"resume prior session", "revive old session", "where was I", "switch to last session", "continue from last session"	recent-sessions.md exists (+1)	—	Sonnet tier; idle-session recovery skill
/cost_snapshot	cost, spend, spent, budget, how much	"how much have I spent", "cost report", "show costs", "project cost", "what has this cost"	—	—	Read-only; safe at any time
/critic	critic, critique, review the plan, find issues, gaps, risks	"critique this plan", "review the plan", "find issues with this plan", "what's wrong with this", "check the plan"	Plan found (+1), no critic-response yet (+1)	Recent commits exist + plan + no review (−2 — prefer /review)	Runs before /implement to catch plan gaps
/discover	discover, scan, map, repos, inventory, dependencies, index	"scan my repos", "map the codebase", "what repos do I have", "index the project", "save the architecture"	repos-inventory.md missing/stale (+2), cache stale (+1)	—	Run once on setup or when repos change
/end_of_day	end of day, EOD, wrapping up, done for the day, save progress	"wrapping up", "done for the day", "save my progress", "end of day", "EOD"	Today's daily cache exists (+1)	No session file today (−1)	Saves state and consolidates work
/end_of_task	end of task, finalize, push, ship, done, complete, wrap up task	"finalize this", "we're done", "ship it", "task complete", "wrap up this task", "push the branch"	APPROVED review found (+2)	On main/clean (−2), uncommitted changes without APPROVED (−1)	High-impact. Requires /review APPROVED first
/gate	gate, check before, quality check, verify before next step	"check before proceeding", "run the gate", "verify before next step", "quality checkpoint"	Active task folder (+1)	No plan, no commits (−1)	Runs between phases; usually auto-invoked by workflow
/implement	implement, code, write code, build, start coding	"implementing a plan", "writing code from a plan", "start coding", "build this", "write the code", "let's code"	Plan found (+2), uncommitted changes (+1), active in_progress session (+1)	No plan found (−2)	High-impact. Requires explicit user command
/init_workflow	init, initialize, bootstrap, set up workflow, first time	"initialize workflow", "set up dev workflow", "install workflow", "bootstrap workflow", "first time setup"	No .workflow_artifacts/ (+2)	Active task folder exists (−2), active session (−2)	Run once per project
/plan	plan, break down, how to implement, task breakdown	"plan this", "create a plan", "break this down", "how should I implement", "implementation plan"	architecture.md exists (+1)	Plan already found at any depth (−1)	Single-pass plan without critic loop
/review	review, verify, check implementation, code review, does this look right	"review my changes", "check the implementation", "verify implementation", "does this look right", "code review"	Plan + recent commits + no review file (+2), active in_progress session (+1)	No recent commits + no plan (−2)	Runs after /implement, before /end_of_task
/revise	revise, fix the plan, address feedback, update the plan	"fix the plan", "address the critic's comments", "update the plan based on feedback", "revise the plan"	critic-response-*.md exists (+2)	No critic-response exists (−2)	Used after /critic; Opus model
/rollback	rollback, undo, revert, go back, undo task	"undo the implementation", "revert the last changes", "go back to before implement", "undo task N", "reset to pre-implementation"	Uncommitted/recent commits exist (+1)	On main/clean (−1)	High-impact. Destructive — reverts commits
/run	run, full pipeline, everything, end to end, full workflow	"run the full workflow", "end to end", "do everything", "full pipeline"	On main/clean (+1)	Any uncommitted work (−1), no plan (−1)	High-impact. Full pipeline from discover to end_of_task
/start_of_day	start of day, SOD, morning, resume, what was I working on	"what was I working on", "resume", "pick up where I left off", "morning standup", "start of day", "SOD"	No daily cache today + no session file today (+2)	Today's daily cache exists (−1)	Morning context restore
/thorough_plan	plan, thorough plan, plan thoroughly, planning cycle	"plan this thoroughly", "detailed plan with review", "plan and critique", "full planning cycle", "plan this"	No plan exists (+1), on main/clean (+1)	Plan already found (−1)	Standard planning entry point; auto-triages Small/Medium/Large
/weekly_review	weekly, week, recap, friday, weekly summary, week review	"weekly summary", "what did I do this week", "week recap", "friday review", "weekly standup", "weekly report"	It's Friday (+1)	Recent commits today only (−1 — likely mid-task)	Run on Fridays or for a week-level summary

---

Section 7: Cross-Skill Disambiguation Table

When keyword collisions produce a tie, use state signals to break it. These rules layer on top of the catalog scores.

Keyword(s)	Competing skills	Disambiguator
"plan", "break down"	/plan, /thorough_plan, /revise, /critic	No plan exists → boost /thorough_plan +1, suppress /revise −2, suppress /critic −2. Plan exists → suppress /thorough_plan −1; boost /revise +1 if critic-response-*.md exists.
"review", "check", "look at"	/review, /critic, /weekly_review	Plan + recent commits → boost /review +2, suppress /critic −2, suppress /weekly_review −2. No recent commits, no plan → boost /critic +1. Day is Friday → boost /weekly_review +1.
"implement", "code", "write"	/implement, /run	No plan found → suppress /implement −2, boost /thorough_plan +1. Plan found + no review → boost /implement +1.
"commit", "ship", "push", "done", "finish"	/end_of_task, /implement	APPROVED review found → boost /end_of_task +2, suppress /implement −1. No review yet → suppress /end_of_task −2, boost /implement +1.
"start", "begin", "morning", "resume"	/start_of_day, /thorough_plan, /init_workflow	No daily cache today → boost /start_of_day +2. Task folder exists → suppress /init_workflow −2.
"run", "everything", "full pipeline"	/run, /thorough_plan	Any uncommitted work or no plan → suppress /run −1. User says "full" or "end to end" → boost /run +1.

---

Section 8: Invocation Policy

All 19 routable skills are propose-only. After the user confirms, /triage prints the command to type and stops. No Skill tool call is ever made.

Rationale:

• No Haiku skill in this workflow uses the Skill tool. Only Opus orchestrators (/run, /thorough_plan) invoke other skills via Skill tool dispatch. Running a Skill tool call from a Haiku model is unverified and potentially unsupported.
• Universal propose-only eliminates this uncertainty entirely. It also means no exception clauses are needed in any target skill's SKILL.md (per lessons 2026-04-13 and 2026-04-14: when a skill adds exception clauses, the target skill's SKILL.md must be updated too — universal propose-only sidesteps this entirely).
• User friction is minimal: the user sees the proposed command and types it.

Skill	Impact level	Rationale for propose-only
/architect	Medium	Universal propose-only. User types /architect.
/capture_insight	Low	Universal propose-only.
/continue_work	Low	Universal propose-only; recovery skill.
/cost_snapshot	Low	Universal propose-only; read-only.
/critic	Medium	Universal propose-only.
/discover	Medium	Universal propose-only.
/end_of_day	Low	Universal propose-only.
/end_of_task	High	Hard rule in CLAUDE.md — explicit invocation only. Universal policy reinforces this. Per lessons 2026-04-13/2026-04-14.
/gate	Medium	Universal propose-only; additionally a context-sensitive checkpoint — user should consciously invoke.
/implement	High	Hard rule in CLAUDE.md — explicit invocation only. Universal policy reinforces this. Per lessons 2026-04-13/2026-04-14.
/init_workflow	Low	Universal propose-only.
/plan	Medium	Universal propose-only.
/review	Medium	Universal propose-only.
/revise	Medium	Universal propose-only.
/rollback	High	Destructive — reverts commits. Universal policy reinforces safety.
/run	High	Full pipeline; huge scope. Universal policy reinforces deliberate invocation.
/start_of_day	Low	Universal propose-only.
/thorough_plan	Medium	Universal propose-only. Standard planning entry point.
/weekly_review	Low	Universal propose-only.

/revise-fast is NOT in this table — it is not a user-facing routing target.

---

Section 9: Ambiguity Handling

Single-candidate flow

Proposed skill: /name
Rationale: <one sentence citing the top 1-2 scoring signals>

→ Type /name to proceed, or say "different skill" to see alternatives.

Reply handling:

• "different skill" → re-enter with the next-highest candidate.
• Any reply containing a /<cmd> token (e.g., "/plan") → treat as Signal A short-circuit on that new command; re-classify.
• Affirmative reply (case-insensitive, ≤ 5 words, no /<cmd> token): "yes", "go", "do it", "ok", "sure", "proceed", "yep", "sounds good", "yeah", "correct" → print → Type /name to proceed. and stop.
• Negation reply: "no", "cancel", "stop", "nope", "nevermind" → stop. Print: "Would you like to describe what you're trying to do, or type a specific /command?"

Propose-only output (always): After any affirmation, print → Type /skill-name to proceed. Never call the Skill tool.

Ambiguous-candidate flow

Triggers when top N candidates are within 1 point of each other, N ∈ {2, 3}.

Multiple candidates match your request. Ranked:

1) /first  — <rationale + top signal>
2) /second — <rationale + top signal>
3) /third  — <rationale + top signal>  (only if N=3)

Which would you like? (type a number, describe more, or type the command directly)

Reply handling:

• "1", "2", or "3" → proceed as single-candidate flow for that skill.
• Reply containing /<cmd> → Signal A short-circuit; re-classify.
• Free-text → re-classify with new text appended. If still ambiguous, ask ONE clarifying question (hard cap):

  "Are you planning new work, reviewing existing work, or wrapping up a completed task?"

• If the response to that one question is still ambiguous or vague → hard exit: "I couldn't determine which skill fits — please type the command directly or run /help to see available skills." No further iteration.

Decline flow

Triggers when no skill scores above 1, OR N > 3 candidates tied, OR top candidates all have High impact with no disambiguating state signal.

I'm not sure which skill fits. Could you clarify:

- Are you planning, implementing, reviewing, or wrapping up?
- Is this a new task or continuation of an existing one?

One question cycle only. If still ambiguous → hard exit: "Please type the command directly, or run /help to see available skills."

---

Section 10: Edge Cases & Guardrails

Explicit-command short-circuit (Signal A): If the scoring prompt (after trigger-phrase stripping) contains a literal /<skill-name> token, bypass all other scoring. Propose that skill immediately with score 3. Exception: /revise-fast → substitute /revise with note.

Multi-stage plan detection: The plan existence signal uses recursive find (maxdepth 4) — not flat file existence. This correctly handles the stage-subfolder layout from lesson 2026-04-18 (e.g., .workflow_artifacts/memory-cache/stage-2/current-plan.md).

Multi-task-folder handling: When multiple active task folders exist and the user's prompt does not name one, state signals read across all folders (e.g., "any plan found" fires if any folder has a plan). Cost recording is skipped silently.

Trigger-phrase stripping: The stripping in Step 1 is case-insensitive and trimmed. Only exact phrase matches are stripped — not substrings. After stripping, normalize whitespace.

revise-fast hidden: Never appear in ranked lists, ambiguous candidate lists, or proposals. If Signal A matches /revise-fast, substitute /revise with the note: "/revise-fast is an internal variant; proposing /revise instead."

Error-tolerant state checks: Every shell command in Step 2 appends 2>/dev/null. Failed checks contribute 0 points. /triage never errors out due to a missing file or failed command.

One-clarification-question hard cap: Across all flows (single, ambiguous, decline), /triage asks at most one clarifying question per session. If still ambiguous after that one question, always hard exit. No loops.

/triage on itself: If the user types /triage as their prompt, Signal A fires immediately and proposes /triage (a self-referential loop). Handle this gracefully: print "You're already running /triage. Please describe what you're trying to accomplish so I can route you to the right skill."

---

Section 11: Important Behaviors

• Never call the Skill tool. This is a hard constraint, not a preference. Universal propose-only for all 18 skills.
• Scoring prompt is stripped input. Always score the trigger-phrase-stripped, whitespace-normalized prompt — never the raw input.
• State signals are evaluated once at bootstrap. Do not re-evaluate mid-conversation.
• /revise-fast is invisible to users. Never surface it in any output.
• When adding a new skill to the workflow, append a row to the catalog in Section 6. The catalog is the routing source of truth — without a row, the skill is unroutable via /triage.
• /triage is not a gate. It does not block phase transitions. It is a convenience router, not a checkpoint.
• Propose-only output is always the same format: → Type /skill-name to proceed. Consistency helps users trust the output and know they must type the command.