@etiquekit/etq 1.0.1 → 1.0.4

package/docs/SEAT_DISCIPLINE.md

@@ -1,19 +1,79 @@
 # Seat Discipline
 
 Seat discipline keeps agent work useful, attributable, and portable across
-runtimes.
+runtimes. Mechanics say what to do; discipline says when to distrust yourself.
+It is organized by moment of need, not by topic.
+
+## On Wake
+
+- Orient before acting: `whereami`, then the task envelope, then the runbook.
+- Read tiered: status line first, then the exact entry, then the full
+  artifact. Never bulk-load directories for "context" — load the specific
+  receipt you were pointed to.
+
+## Before Claiming Anything
+
+- Verify at claim time: re-read the source at the moment you assert it,
+  never from memory of an earlier read. A cited manifest or spec is read in
+  the same act as the claim.
+- Check the live tip, not your notes: logs and board notes lag integrations.
+- When another seat or a human holds the ground truth, ask — do not
+  solo-reconstruct their state.
+
+## Before Writing
+
+- Read the machine fields, not the title: the envelope's mode and allowed
+  writes decide what a task may do; prose is not authority.
+- Search the exact target surface before classifying work as new — most
+  "add X" tasks are "finish the X that half-exists."
+- One writer per canonical mutation. Design in parallel; mutate single-file.
+
+## While Working
+
+- Technical register on the record: schemas, fields, commands, refs.
+  Metaphor and personality stay in chat; the record is for the next reader.
+- No person or vendor names in artifacts. Attribute external material at the
+  moment of incorporation, with license.
+- One meaningful update per note, reader-resolvable without tooling. Write
+  runbook notes, not raw transcript dumps.
+
+## Before Returning
+
+- Receipt first: outcome, changed files, validation command and result,
+  blockers, next owner. No raw logs — cite paths to them.
+- Run validation before returning the receipt. A red validation is reported
+  red; "done" without proof is a defect, not a return.
+- If your change altered a source shape, refresh its recorded evidence in
+  the same commit.
+- Run post-commit hygiene (`git diff --check HEAD~..HEAD`) — pre-commit
+  checks miss freshly added files.
+
+## On Review
+
+- Verify the shipped artifact, not the branch state you remember.
+- Reviews return evidence, never authority: confirm, counter, or advise.
+  A reviewer packet does not mutate the lane it reviews.
+- Accepted review is different from integrated work.
+- Proposals carry their own counter-arguments and open questions — invite
+  the pressure test.
+
+## At Authority Edges
+
+- Stop when authority is unclear, validation cannot run, writes would exceed
+  the envelope, a secret appears, or the session is unbounded.
+- Never claim work was authorized without decision language from the
+  authority holder. Enthusiasm is not routing.
+- Risk-tier your own gates: proceed on cheap-and-reversible; insist on a
+  real decision for irreversible, authority-changing, or spend actions.
+  Gating everything trains the rubber-stamp.
+- Promote only through the policy-resolved gate.
+- Memory proposes. Ledgers authorize. Receipts prove. Audit accounts.
 
-## Operating Rules
+## When Blocked
 
-- Read the repo state before forming a plan.
-- Use a task envelope before product edits.
-- Keep each session bounded.
-- Write useful runbook notes, not raw transcript dumps.
-- Run validation before returning a receipt.
-- Keep receipts compact and cited.
-- Treat accepted review as different from integrated work.
-- Promote only through the policy-resolved gate.
-- Stop when authority, secrets, scope, or validation are unclear.
+- Use the diagnostics verbs before hand-debugging; walk the recovery path
+  from the user's seat before declaring a state recoverable.
+- Escalate with a concrete blocker, not a vibe. Quiet is not a finding.
 
 ## Identity Split
 
@@ -27,42 +87,30 @@ repo_substrate: source, ledger, receipts, validation, commits
 
 ## Work Boundaries
 
-A seat should know:
-
-- allowed writes;
-- forbidden writes;
-- validation commands;
-- stop conditions;
-- expected receipt;
-- next owner.
-
-If any of those are missing, ask for a sharper task.
-
-## Session Hygiene
-
-Sessions are for bounded work. Restoring a session restores context, not
-authority.
-
-## Review Discipline
-
-Review seats should pressure-test hidden authority changes, missing validation,
-oversized write sets, stale assumptions, unbounded sessions, evidence that
-cannot be reproduced, and claims that should be receipts.
-
-Reviewers contribute evidence. The routed owner mutates truth.
+A seat should know: allowed writes; forbidden writes; validation commands;
+stop conditions; expected receipt; next owner. If any are missing, ask for a
+sharper task. Sessions are for bounded work; restoring a session restores
+context, not authority.
 
 ## Tool Discipline
 
-Use chat for conversation, planning tools for intent, code review tools for
-diffs and CI, incident tools for runtime evidence, and identity providers for
-authentication.
-
-Do not let any of them become acceptance.
+Chat is conversation, planning tools are intent, code review tools are diffs
+and CI, incident tools are runtime evidence, identity providers are
+authentication. None of them are acceptance.
 
 ## Failure Signals
 
 Open a follow-up or stop the run when a seat edits outside allowed writes,
-validation repeatedly fails, the runbook grows without a receipt, a task spans
-unrelated repos, a runtime claims success without evidence, a projection
-disagrees with git, a grant owner is unclear, or an integration tries to close
-or promote work.
+validation repeatedly fails, the runbook grows without a receipt, a task
+spans unrelated repos, a runtime claims success without evidence, a
+projection disagrees with git, a grant owner is unclear, or an integration
+tries to close or promote work.
+
+## Growing Local Discipline
+
+This document ships the judgment that is true everywhere. Your team's own
+calibration — which seats are strong where, which tools misbehave, which
+routes need a second look — is grown locally: when a correction lands, record
+it as one small note with why and how-to-apply, index it, and re-read it at
+the moment you would otherwise repeat the mistake. Corrections that are never
+recorded are corrections you will pay for twice.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@etiquekit/etq",
-  "version": "1.0.1",
+  "version": "1.0.4",
   "description": "Governance and evidence CLI for AI-assisted engineering seats.",
   "private": false,
   "type": "module",