@sjawhar/opencode-legion-envoy 1.17.1 → 1.18.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -13875,7 +13875,7 @@ var dispatchToolSpecs = [
13875
13875
  },
13876
13876
  {
13877
13877
  name: "dispatch_doc_edit",
13878
- description: "Apply deterministic document edits, including retyping an identified paragraph into a schema-declared typed block. " + "Do not use it for review feedback or for reading; use dispatch_comment, dispatch_suggest, or dispatch_doc_read instead. " + `The spec (or any document) holds requirements, design, and decisions - never progress, status, or timestamps. ${OWNER_REFERENCE} ${SPEC_WRITING_GUIDANCE}`,
13878
+ description: "Apply deterministic document edits, including retyping an identified paragraph into a schema-declared typed block. " + "Do not use it for review feedback or for reading; use dispatch_comment, dispatch_suggest, or dispatch_doc_read instead. " + "For replace, delete, and quote insert anchors, find text exactly as rendered: omit Markdown markers such as backticks or asterisks. " + `The spec (or any document) holds requirements, design, and decisions - never progress, status, or timestamps. ${OWNER_REFERENCE} ${SPEC_WRITING_GUIDANCE}`,
13879
13879
  arguments: (z) => ({
13880
13880
  issue: z.string().describe(ISSUE_REFERENCE).optional(),
13881
13881
  project: z.string().describe("Project key owning the document.").optional(),
@@ -14165,8 +14165,9 @@ var stateTree = strictObject({
14165
14165
  locator: stateTreeLocator.optional()
14166
14166
  });
14167
14167
  var stateGate = strictObject({
14168
- designAskId: nonEmptyString.optional(),
14169
- designApproved: nonEmptyString.optional()
14168
+ artifactId: nonEmptyString,
14169
+ latestVersion: number2().int().positive(),
14170
+ approvedVersion: number2().int().positive().optional()
14170
14171
  });
14171
14172
  var stateRole = strictObject({
14172
14173
  role: nonEmptyString,
@@ -14308,7 +14309,11 @@ var LegionDaemonApi = {
14308
14309
  response: object({})
14309
14310
  },
14310
14311
  GatesRegister: {
14311
- request: architectCapability.extend({ issue: nonEmptyString, askId: nonEmptyString }),
14312
+ request: architectCapability.extend({
14313
+ issue: nonEmptyString,
14314
+ artifactId: uuid2(),
14315
+ version: number2().int().positive()
14316
+ }),
14312
14317
  response: object({})
14313
14318
  },
14314
14319
  Grant: {
@@ -15649,7 +15654,7 @@ ${trailer.join(`
15649
15654
  const result = await client.requestApproval(resolved.artifact.id, { actor });
15650
15655
  if (result.ask === null) {
15651
15656
  return {
15652
- text: `${resolved.artifact.name} is already approved at version ${result.version} by ${result.approval.by?.id ?? "unknown"}; no new request was opened. An edit after approval makes it stale, so request again only for a new version.`,
15657
+ text: `${resolved.artifact.name} (document id ${resolved.artifact.id}) is already approved at version ${result.version} by ${result.approval.by?.id ?? "unknown"}; no new request was opened. An edit after approval makes it stale, so request again only for a new version.`,
15653
15658
  details: {
15654
15659
  ...resolved.owner.kind === "project" ? documentResultDetails(resolved.artifact) : { issue: resolved.issue?.key },
15655
15660
  artifact: resolved.artifact.id,
@@ -15659,7 +15664,7 @@ ${trailer.join(`
15659
15664
  }
15660
15665
  const details = await askResultDetails(client, result.ask, resolved);
15661
15666
  return {
15662
- text: `Approval requested for ${resolved.artifact.name} at version ${result.version} (ask ${result.ask.id}). The answer arrives as artifact.approved or artifact.changes_requested; an edit after approval makes it stale, so request again for the new version.`,
15667
+ text: `Approval requested for ${resolved.artifact.name} (document id ${resolved.artifact.id}) at version ${result.version} (ask ${result.ask.id}). The answer arrives as artifact.approved or artifact.changes_requested; an edit after approval makes it stale, so request again for the new version.`,
15663
15668
  details: { ...details, artifact: resolved.artifact.id, version: result.version }
15664
15669
  };
15665
15670
  }
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@sjawhar/opencode-legion-envoy",
3
- "version": "1.17.1",
3
+ "version": "1.18.1",
4
4
  "type": "module",
5
5
  "main": "dist/src/server.js",
6
6
  "exports": {
package/skills/AGENTS.md CHANGED
@@ -34,8 +34,11 @@ durable human question; replies come back to the worker's own session.
34
34
  ## Durable artifacts
35
35
 
36
36
  Phase handoffs are committed in lifecycle order: architect, plan, implement, test, and review.
37
+ Only the implementer pushes them: it and the merger act as the code-writing GitHub App, while
38
+ the planner, tester, reviewer, and architects act as the review App (`appRoleForLegionRole`,
39
+ `packages/daemon/src/daemon/github-apps.ts`), which holds no `contents` permission — their
40
+ handoff commits stay on the shared workspace's issue branch and ride the implementer's next push.
37
41
  A clean review ends with the `.legion/` deletion pushed by the implementer at the reviewer's
38
- direction (the review App holds no `contents` permission and cannot push), which the reviewer
39
- then approves; retro records its learning in
42
+ direction, which the reviewer then approves; retro records its learning in
40
43
  `docs/solutions/` and writes no handoff. GitHub comments and reviews carry the required Legion
41
44
  footer so the daemon can attribute artifacts to their worker session.
@@ -243,9 +243,10 @@ type EditOp = {
243
243
  };
244
244
  ```
245
245
 
246
- Target `replace` and `delete` by the document's plain text: inline-code and link text match without Markdown syntax, and a table-cell
247
- anchor is its cell text. Quote code-block contents without their Markdown fences. A quote must stay within one textblock; split changes
248
- that span separate blocks into separate operations.
246
+ Target `replace`, `delete`, and quote insert anchors by a block's text as rendered: write inline
247
+ code without backticks, bold without asterisks, and link text without link syntax. A table-cell
248
+ anchor is its cell text. Quote code-block contents without their Markdown fences. A quote must stay
249
+ within one textblock; split changes that span separate blocks into separate operations.
249
250
 
250
251
  `replace` requires `find` and `with`; `delete` requires `find`; `insert` requires `markdown` and exactly one of `after` or `before`. An
251
252
  insert anchor is a quote, `"start"`, `"end"`, or `"heading:Title"`. Ordinary inserts create a sibling block before or after the quote or
@@ -286,6 +287,11 @@ Should we ship the migration?
286
287
  :::
287
288
  ```
288
289
 
290
+ When a human answers a decision written as an ask block, the answer lives on that ask. Use
291
+ `dispatch_resolve_ask` when the decision is resolved without a human response, or preserve the
292
+ human's answer; never rewrite the question into its answer or blank its options. An edit that leaves
293
+ an ask block without a question or with a blank option is rejected with `INVALID_ASK_BLOCK`.
294
+
289
295
  ## Comments and suggestions
290
296
 
291
297
  Add feedback with:
@@ -18,7 +18,11 @@ separate coordinator to finish necessary work.
18
18
  Message a known phase worker with `envoy_publish` to `notifications.role.` followed by
19
19
  its encoded role token (the token `spawn_worker` returned for it); re-assign it by
20
20
  calling `spawn_worker` again on the same existing role, which resumes the same process
21
- instead of starting a fresh one. Phase workers escalate lifecycle, scope, and
21
+ instead of starting a fresh one. Phases on one issue are strictly sequential -- one role
22
+ is the issue's active phase at a time, and calling `spawn_worker` for a different role
23
+ while a phase is active supersedes that phase: the superseded worker's
24
+ `legion handoff complete` is then refused, so finish (or deliberately abandon) one role
25
+ before assigning the next. Phase workers escalate lifecycle, scope, and
22
26
  cross-phase matters the same way: `envoy_publish` to your own encoded token. Any role
23
27
  may use `dispatch_ask` directly for a standalone human question; replies return to the
24
28
  asking session.
@@ -29,7 +33,8 @@ separate coordinator to finish necessary work.
29
33
  hold (your own, or one `spawn_worker` returned) or compute another with the
30
34
  `roleToken` helper from `@legion/contracts` exactly the way the daemon does.
31
35
  - There is no label vocabulary. Dispatch status replaces the board, and the design gate
32
- is a `dispatch_ask` answered `Approve`, not a label. Never attempt to apply a label.
36
+ is a human approving the root spec document at a version in Dispatch, requested with
37
+ `dispatch_request_approval` — not a label and not an ask. Never attempt to apply a label.
33
38
  - Deferring necessary work is failure. The sole valid deferral is a new child issue you
34
39
  create and continue to own. Re-file a genuinely independent child through the
35
40
  controller rather than treating it as an abandoned dependency.
@@ -83,34 +88,51 @@ adoption/decomposition and waves, acceptance criteria, and the integration test
83
88
  "spec" artifact beside it (`dispatch_artifact` with the primary document's name replaces the
84
89
  human's document; do not do that). Both readers described in
85
90
  [Writing for the human](../dispatch/SKILL.md#writing-for-the-human) must be able to follow it.
86
- When the config-armed root design gate applies, run this exact sequence **before any
87
- Legion-role spawn**, including a sub-architect:
91
+ The design gate runs only when the "Design gate policy" line at the end of your system prompt
92
+ says `gates.design: root-issues`. When it says `gates.design: off`, write the spec and continue
93
+ to section 2 with no approval step at all: do not request approval, do not register a gate, and
94
+ do not wait for `design-approved`. A sub-architect on a child issue has no policy line and never
95
+ runs the gate either: the root approval covers the tree. When the gate is armed, run this exact
96
+ sequence **before any Legion-role spawn**, including a sub-architect:
88
97
 
89
98
  ```text
90
99
  dispatch_doc_edit({ issue: "<root issue>", ... }) // extend the primary document in place
91
- askId = dispatch_ask({
100
+ result = dispatch_request_approval({ issue: "<root issue>" }) // the primary document by default
101
+ legion({
102
+ op: "register_gate",
92
103
  issue: "<root issue>",
93
- question: "<what is true today, in one sentence> <what will be true when this lands, in one sentence> <how: one issue or N child issues, and what the first step is> I recommend Approve because <one reason>.",
94
- options: [
95
- { label: "Approve", description: "Work starts as described; the first worker is spawned now." },
96
- { label: "Hold", description: "Nothing starts; reply on the issue with what should change first." },
97
- ]
104
+ artifactId: result.details.artifact, // the document id, a UUID such as 4e0aca36-77b3-43bd-96cf-d58890ae64e4
105
+ version: result.details.version, // the version number the human is asked to approve
98
106
  })
99
- legion({ op: "register_gate", issue: "<root issue>", askId })
100
107
  ```
101
108
 
102
- Both options are required: a question with only `Approve` is not a decision. The whole ask is
103
- read on a phone by someone who has not read the code: no file paths, line numbers, document
104
- versions, or role tokens in it. Sami, 2026-09-12, on a gate ask that broke this rule: "I have no
105
- idea what the fuck you're talking about."
106
-
107
- Then park. Do not release a wave or spawn a Legion role until a later delivered wake
108
- shows `design-approved` on the root. On a deployment whose design gate is off
109
- (`gates.design: off` in its `legion.yaml`), the daemon satisfies the gate as you register it
110
- and `design-approved` arrives immediately proceed. The daemon then closes the ask on Dispatch
111
- (you will see `ask.resolved` for it); that is expected and needs nothing from you. Approval
112
- covers the entire tree: later waves, re-scopes, and integration-failure children do not repeat
113
- this sequence.
109
+ `dispatch_request_approval` opens a system question on the document with the fixed options
110
+ `Approve` and `Request changes`; a human answers it from the Inbox or approves from the
111
+ document's own header. Never open a `dispatch_ask` with an `Approve` option yourself: an
112
+ ordinary question is not a gate and the daemon ignores its answer. Copy `artifactId` and
113
+ `version` from the result of `dispatch_request_approval` — its text reads "Approval requested for
114
+ spec.md (document id <UUID>) at version <N>" and its `details.artifact` / `details.version` carry
115
+ the same two values. The document id is never the slug or file name you passed in (`spec`,
116
+ `spec.md`): the daemon recognizes the document's approval events by that id, and both the
117
+ `legion` tool and the daemon refuse a value that is not a UUID. Calling
118
+ `dispatch_request_approval` again while a request is open returns the same open request, so it
119
+ is safe to repeat. If its text instead reads "spec.md (document id <UUID>) is already approved at
120
+ version <N>" — a human approved from the document header before you asked — still call
121
+ `register_gate` with that id and version: the daemon reads the approval from Dispatch as it
122
+ registers, opens the gate, and delivers `design-approved` at once. The same read covers a human
123
+ who answers the question between your `dispatch_request_approval` and `register_gate` calls, so
124
+ an approval is never lost to timing; you never approve anything yourself.
125
+
126
+ Then park. Do not release a wave or spawn a Legion role until a later delivered wake shows
127
+ `design-approved` on the root. On `design-changes-requested`, revise the spec (a new version of
128
+ the primary document), call `dispatch_request_approval` again — it re-opens the request at the
129
+ new version — and stay parked. Approval is pinned to the spec version: editing the root spec
130
+ after approval closes the gate again with no wake (you made the edit, or the `artifact.version`
131
+ event on your issue tells you), so call `dispatch_request_approval` again, and release no new
132
+ wave and spawn no new role until the next `design-approved` arrives — work already in flight
133
+ continues. Later waves, re-scopes, and integration-failure children that leave the root spec
134
+ untouched need no new approval, and a child issue's spec is never gated: the root approval covers
135
+ the tree.
114
136
 
115
137
  ## 2. Children in flight
116
138
 
@@ -141,6 +163,12 @@ in flight. On each child closure, re-scope open work, close obsolete work with a
141
163
  release the next wave only when it now makes sense. There is no inter-child dependency
142
164
  mechanism to encode.
143
165
 
166
+ Release admits nothing. A child never takes an admission slot or becomes a root tree of its
167
+ own: the daemon ignores a child's `todo` while your tree is live, and this `spawn_worker` is
168
+ what starts the child — the daemon writes its Dispatch status `in_progress` on the first
169
+ sub-architect spawn while the child is at `todo`. A released child with no sub-architect stays
170
+ at `todo` until you spawn one.
171
+
144
172
  ## 3. Children complete
145
173
 
146
174
  Treat `children-complete` as the edge into the end-game, not as a reason to close the
@@ -182,13 +210,13 @@ legion({
182
210
  op: "spawn_worker",
183
211
  issue: "LEGION-40",
184
212
  role: "implementer",
185
- task: "Run the legion-retro skill now. Capture durable learnings and post the issue comment; do not create a .legion handoff file."
213
+ task: "Run the legion-retro skill now. Capture durable learnings and post the retro message on the Dispatch issue with dispatch_message; do not create a .legion handoff file."
186
214
  })
187
215
  ```
188
216
 
189
217
  Wait for the implementer to report its durable retro result. Retro output is
190
- `docs/solutions/` plus an issue comment; it must not create a `.legion` file or change
191
- the reviewer-approved head after cleanup.
218
+ `docs/solutions/` plus one `dispatch_message` on the issue; it must not create a `.legion`
219
+ file or rewrite the reviewer-approved head after cleanup.
192
220
 
193
221
  ## 6. Architect sign-off and merge
194
222
 
@@ -202,19 +230,37 @@ Preserve this order exactly:
202
230
  2. on a clean review, `spawn_worker` the implementer once more to push only the `.legion/`
203
231
  deletion (the review App holds no `contents` permission), then the reviewer approves that
204
232
  head. The deletion must land before that approval, which is head-pinned. An implementer
205
- completion always writes the issue's status as `testing`; this one is not a test round,
206
- so on its `phase-complete` wake call `legion({ op: "set_status", issue, status: "retro" })`
207
- before `spawn_worker` on the reviewer to approve that head (a finished reviewer may already
208
- be retired; `spawn_worker` resumes it);
209
- 3. retro completes without dirtying the branch beyond `docs/solutions/`;
210
- 4. the merger verifies the current head is the reviewer-approved head plus only the retro
211
- commits and publishes `READY #<n> at <sha>` to `notifications.role.pr-queue`; it never
233
+ completion advances the status only from `in_progress` to `testing`; this push, like retro
234
+ later, leaves the status where it is, so you set nothing by hand — on its `phase-complete`
235
+ wake, `spawn_worker` the reviewer to approve that head (a finished reviewer may already be
236
+ retired; `spawn_worker` resumes it);
237
+ 3. retro commits its learnings under `docs/solutions/` on top of the approved head; that
238
+ commit does not void the approval and never returns the tree to the tester or reviewer;
239
+ 4. the merger verifies the current head is the reviewer-approved head plus only commits that
240
+ change `docs/solutions/` (`jj diff --from <approved-sha> --to <tip-sha> --summary`, quoted in READY)
241
+ and publishes `READY #<n> at <sha>` to `notifications.role.pr-queue`; it never
212
242
  merges. The merge queue merges under its own authority and the repository's own rules
213
243
  (branch protection, CODEOWNERS); whether a human must approve first is that repository's
214
244
  setting, not Legion's, and you never ask for or wait on such an approval.
215
245
 
216
- If anything else changes the head, return to review; do not let the merger publish `READY`
217
- for an obsolete approval.
246
+ What returns the tree to review: a changed diff — a commit above the approved head that
247
+ touches anything outside `docs/solutions/`, or a rebase whose fingerprint (the `legion-worker`
248
+ skill's unchanged-diff check) differs from the approved head's. What does not: retro's
249
+ `docs/solutions/` commit, and a rebase forced by a GitHub-reported conflict whose fingerprint
250
+ is unchanged. For that rebase the order is: the implementer rebases and posts the before/after
251
+ fingerprints; the tester re-runs the bare gates only; the reviewer confirms and approves the new
252
+ head by SHA (or continues its round if it had not approved); the merger republishes READY.
253
+ Retro does not re-run. A rebase happens only when GitHub reports `CONFLICTING`
254
+ (`legion gh -- pr view <n> --json mergeable,mergeStateStatus`); read that on every end-game
255
+ wake — `pr-ready`, `pr-review`, `phase-complete`, `catchup-overseer` — because a `CONFLICTING`
256
+ PR gets no CI and no wake announces it, and send the implementer to rebase the moment you see
257
+ it. Do not let the merger publish `READY` for an obsolete approval.
258
+
259
+ If a worker reports that `legion threads resolve` exited 1 naming a review thread GitHub refused
260
+ to resolve, open an action ask (`dispatch_ask` with `kind: "action"`) that names the thread's URL
261
+ and GitHub's message for a human to resolve it by hand; the merger does not publish while it is
262
+ open. That is the one review-thread step a human takes: the review App cannot resolve threads,
263
+ and the implementer's and merger's runs of the command close every accepted one.
218
264
 
219
265
  ## 7. Close
220
266
 
@@ -236,19 +282,23 @@ corresponding lifecycle procedure.
236
282
 
237
283
  | Wake | Procedure |
238
284
  | --- | --- |
285
+ | `child-adopted` | Payload `{type:"child-adopted", child, remaining}`. A child is now in your tree — one created under this issue (by you or a human), or one a daemon upgrade moved back into your tree from a root tree of its own (LEGION-57). If Dispatch shows it released **and open** — `todo` through `retro`, never `done`; `remaining` counts exactly those — and `legion state` shows no `roles` entry with `issue` = the child and `role: "architect"`, `spawn_worker` its architect now. An unreleased child waits for its wave; a `done` child is finished and gets nothing, whatever stray tree of its own `legion state` may still show. |
286
+ | `child-status` | Payload `{type:"child-status", child, from, to}`. Your child's Dispatch status changed. `to: "todo"` with no architect claim for the child (`legion state`) means it is released and unowned — your own `release_wave` echo, or a human's move — so `spawn_worker` its architect. `to: "backlog"` or `"icebox"` means the child was de-prioritised (a human's move, or your own `set_status`): a child has no tree of its own, so the daemon stops nothing on that move — tell its sub-architect (`envoy_publish` to its role topic) to finish the step in flight and park, or re-scope it; its finished workers idle-retire, and it resumes from its session on your next `spawn_worker` once the child is released again. Any other transition is information for re-scoping. |
239
287
  | `child-closed` | Read the child completion and remaining open children. Re-scope or close obsolete open work; release an appropriate next wave, or await `children-complete`. |
240
288
  | `children-complete` | Execute steps 3–4: parent integration verification; failures become a new child wave, success advances to review and retro. |
241
289
  | `child-reopened` | Treat the completion edge as reset. Reassess the reopened child and return the tree to children-in-flight; do not continue an already-started end-game. |
242
- | `phase-complete` | Payload `{type:"phase-complete", issue, role, summary}`. May arrive live or via `catchup-overseer`'s `phaseCompletions`. Read the committed handoff for that phase, then spawn the next phase's owner, or `spawn_worker` on the same role again to resume it with corrections if the handoff shows unresolved gaps. A `reviewer` completion whose GitHub review is `CHANGES_REQUESTED` (the daemon has already returned the issue's Dispatch status to `in_progress` for this) means `spawn_worker` the **implementer** again with the review findings — thread URLs and blocking items — as its task, then route back through tester and reviewer in order; never `spawn_worker` the reviewer directly off this wake and never proceed to retro on this verdict. A reviewer completion with an `APPROVED` review proceeds to retro (step 5). |
243
- | `worker-queued` | Payload `{type:"worker-queued", issue, role}`. The deployment's worker cap is full; this role's spawn is queued. Do not respawn or retry wait for `worker-started`. |
290
+ | `design-approved` | Payload `{type:"design-approved"}`. A human approved the root spec document at its current version; the gate is open. Proceed to section 2. |
291
+ | `design-changes-requested` | Payload `{type:"design-changes-requested", version, reason, author?}`. A human asked for changes to the root spec at `version`, for `reason`. Revise the spec, call `dispatch_request_approval` again, and stay parked; the gate is closed. |
292
+ | `phase-complete` | Payload `{type:"phase-complete", issue, role, summary}`. May arrive live or via `catchup-overseer`'s `phaseCompletions`. Read the committed handoff for that phase, then spawn the next phase's owner, or `spawn_worker` on the same role again to resume it with corrections if the handoff shows unresolved gaps. A `reviewer` completion whose GitHub review is `CHANGES_REQUESTED` (the daemon returns the issue's Dispatch status to `in_progress` for this, on the reviewer's completion and again when you spawn the corrective implementer unless the daemon already knows the issue is `in_progress`) means `spawn_worker` the **implementer** again with the review findings — thread URLs and blocking items — as its task, then route back through tester and reviewer in order; never `spawn_worker` the reviewer directly off this wake and never proceed to retro on this verdict. A reviewer completion with an `APPROVED` review proceeds to retro (step 5). A `reviewer` completion after a conflict-forced rebase whose review body names an unchanged fingerprint is a confirmation, not a round: if retro already completed, `spawn_worker` the merger; otherwise resume the step you were on. |
293
+ | `worker-queued` | Payload `{type:"worker-queued", issue, role}`. This role's task is queued for promotion — either the deployment's worker cap is full, or the live worker acknowledged the task without starting a turn and the daemon is retrying it (counted; the worker is replaced after three such failures, still with the same task). Do not respawn or retry — wait for `worker-started`. |
244
294
  | `worker-started` | Payload `{type:"worker-started", issue, role}`. A previously queued role has been promoted and is now running. Treat it exactly as a normal spawn: resume tracking that role's live session. |
245
295
  | `pr-ready` | Verify the live PR head, green status, and review state. Continue the review/retro/merger order only for that current head. |
246
- | `pr-review` | Payload `{type:"pr-review", state, author, body}`. Delivered to whichever role is currently active for the issue, falling back to you when no worker phase is active. Follows the same verdict rule as a reviewer's `phase-complete`: `state: "changes_requested"` sends the implementer back in with the review findings, then tester, then reviewer — never the reviewer again and never retro; `state: "approved"` proceeds toward retro (step 5) once the step 6 integration/merge-gate conditions are met. |
296
+ | `pr-review` | Payload `{type:"pr-review", state, author, body}`. Delivered to whichever role is currently active for the issue, falling back to you when no worker phase is active. Follows the same verdict rule as a reviewer's `phase-complete`: `state: "changes_requested"` sends the implementer back in with the review findings, then tester, then reviewer — never the reviewer again and never retro; that `spawn_worker` returns the issue to `in_progress` on its own (the daemon writes it for a corrective implementer whenever the PR's latest recorded review is changes requested, a human's after approval included), so you set nothing by hand; `state: "approved"` proceeds toward retro (step 5) once the step 6 integration/merge-gate conditions are met. `state: "approved"` on a rebased head whose body names an unchanged fingerprint is that confirmation: proceed to retro if it has not run, otherwise to the merger — never to a second retro or test round. |
247
297
  | `pr-blocked` | Payload `{type:"pr-blocked", pr, attempts}`. `attempts` counts heads pushed onto a red verdict that changed something outside `.legion/` — handoff-only pushes (`.legion/` paths only) never count; a push the daemon cannot classify (a listener without `changed_paths`, a list capped at 100, a push listing no commits) does. Published once per exhausted count, not on every later red verdict for that count. Read the failed CI evidence and recovery attempts. Assign a focused implementer or corrective child, then return it through testing and review; do not treat the blocked PR as final. |
248
298
  | `pr-merged` | Payload `{type:"pr-merged", pr, mergeCommitSha}`. The merge queue landed the PR. This is your cue for step 7: post the sign-off comment naming that merge commit and set the issue `done`. Nothing else follows a merge. |
249
299
  | `pr-closed-unmerged` | Decide from current scope whether to reopen the work, send a fresh implementer, or cancel it with a reason. Delegate the repository action to the responsible phase worker and keep ownership. |
250
300
  | `issue-comment` | Interpret the comment in the issue's design context. Answer it, adjust the plan, or relay it via `envoy_publish` to the responsible worker's role token; scope and product decisions remain with you. |
251
- | `catchup-overseer` | Verify its gates, child counts, and PR verdicts against current artifacts, then resume the applicable numbered lifecycle step. It is a current-state snapshot, not a raw-event replay. For each entry in its `phaseCompletions` (`{issue, role, summary, at}`, phases that completed while you were not live), handle it exactly as a `phase-complete` wake. |
301
+ | `catchup-overseer` | Verify its gates, child counts, and PR verdicts against current artifacts, then resume the applicable numbered lifecycle step. It is a current-state snapshot, not a raw-event replay. `gates[LEGION_TREE].open` is the design gate's current state: `true` means the root spec is approved at its current version and you may spawn; `false` (or no `open` key, meaning no gate is registered) means the sequence in section 1 still applies. For each entry in its `phaseCompletions` (`{issue, role, summary, at}`, phases that completed while you were not live), handle it exactly as a `phase-complete` wake. Then compare `childCounts[LEGION_ISSUE].open` (the children not `done`) with `legion state` and Dispatch: any **open** released child — `todo` through `retro` — with no architect role claim gets `spawn_worker` for its architect, a `child-adopted` or `child-status` wake you missed while not live; a `done` child gets nothing, whether or not a lingering legacy tree of its own still shows in `legion state`. |
252
302
  | `worker-died` | Payload `{type:"worker-died", issue, role}`. The daemon probed and retried this role's worker through `MAX_LAUNCH_FAILURES` attempts and could not confirm a boot — never a raw-event replay or a silent revive. Reassess the work and `spawn_worker` again for the role (it resumes the same agent via `--resume` if a session file survived) or reassign it if the failure looks environmental, not agent-specific. |
253
303
  | `reopened` | Reopen the root lifecycle: inspect the reason and current artifacts, reassess scope and children, and resume at the first applicable numbered step. |
254
304
 
@@ -129,7 +129,11 @@ on stale entries until their source artifact explains the anomaly.
129
129
  ## Mentions
130
130
 
131
131
  Read the mention and its artifact. Answer it when it asks the controller for triage or
132
- human-facing information. Otherwise resolve the authoritative owning architect role and
133
- route the verified context with `envoy_publish`. Do not route raw event traffic or invent a
134
- role token from a partial issue reference.
132
+ human-facing information. A human asking how to let a root proceed past its design gate
133
+ approves the root issue's spec document in Dispatch the `Approve` control in the document's
134
+ header, or the approval question the architect's request opened in the Inbox. The controller
135
+ never opens a gate and there is no operator command for it; a project that does not want the
136
+ gate at all runs `gates.design: off` in its `legion.yaml`. Otherwise resolve the authoritative
137
+ owning architect role and route the verified context with `envoy_publish`. Do not route raw
138
+ event traffic or invent a role token from a partial issue reference.
135
139
 
@@ -17,15 +17,23 @@ retrospective's durable output.
17
17
  1. Tester green and all code-review cycles finish.
18
18
  2. The implementer pushes the `.legion/` deletion at the reviewer's direction, and the reviewer
19
19
  approves that head.
20
- 3. Run this retro: commit durable learnings to `docs/solutions/` and post the issue comment.
20
+ 3. Run this retro: commit durable learnings to `docs/solutions/` and post the retro message on
21
+ the Dispatch issue.
21
22
  Retro writes **no `.legion` file**, so it never re-dirties the cleaned handoff tree.
22
- 4. The merger verifies the approved head, publishes `READY`, and pushes nothing; the merge queue
23
- merges under the repository's own rules.
23
+ 4. The merger verifies the tip is the approved head plus commits that change only
24
+ `docs/solutions/` `jj diff --from <approved-sha> --to <tip-sha> --summary`, quoted in READY —
25
+ publishes `READY`, and pushes nothing; the merge queue merges under the repository's own
26
+ rules.
24
27
  5. After the merge lands, the implementer — not the reviewer, the merger, or the queue — verifies
25
28
  the change in production and records it on the PR and the issue (Sami, 2026-09-13, verbatim:
26
29
  "the agent that developed it should be responsible for testing in production"). The
27
30
  architect's sign-off waits for that record.
28
31
 
32
+ Retro's commit sits above the reviewer's approved head and the approval stands: a commit that
33
+ changes only `docs/solutions/` does not void it, and the tree goes from retro to the merger —
34
+ never back to the tester or reviewer. A conflict-forced rebase after retro moves these documents
35
+ with the branch; retro does not re-run.
36
+
29
37
  Do not start retro before step 2, skip it because the change seems mechanical, or publish `READY`
30
38
  before step 3. The design gate is not a substitute for review and retro.
31
39
 
@@ -59,18 +67,22 @@ date: YYYY-MM-DD
59
67
  status: active
60
68
  module: affected-module
61
69
  related_issues:
62
- - "owner/repo#123"
70
+ - "LEGION-123" # the Dispatch issue
71
+ - "owner/repo#456" # the pull request
63
72
  ---
64
73
  ```
65
74
 
66
75
  Commit the documentation on the existing issue branch, advance its existing bookmark, and push
67
- that branch. Do not create a replacement branch or bookmark. Then post an issue comment naming
68
- the documents and the one-to-three most useful takeaways. The comment must carry this revived
69
- implementer's structured attribution footer with `phase` set to `retro`:
76
+ that branch. Do not create a replacement branch or bookmark. Then post one Dispatch message on
77
+ the issue `issue` is your `LEGION_ISSUE`; Legion issues live on Dispatch, never on a GitHub
78
+ issue, and the `gh` shim refuses every GitHub-issue write — naming the documents and the
79
+ one-to-three most useful takeaways. The message must carry this revived implementer's structured
80
+ attribution footer with `phase` set to `retro`; the body is capped at 2,000 characters:
70
81
 
71
- ```bash
72
- legion gh -- issue comment <issue-number> \
73
- --body $'## Retro Complete
82
+ ```ts
83
+ dispatch_message({
84
+ issue: "<KEY>",
85
+ body: `## Retro Complete
74
86
 
75
87
  **Learnings documented in:**
76
88
  - docs/solutions/<path>.md
@@ -78,19 +90,22 @@ legion gh -- issue comment <issue-number> \
78
90
  **Key takeaways:**
79
91
  - <reusable lesson>
80
92
 
81
- <!-- legion: {"session":"<session-id>","phase":"retro"} -->' \
82
- --repo <owner>/<repo>
93
+ <!-- legion: {"session":"<session-id>","phase":"retro"} -->`,
94
+ })
83
95
  ```
84
96
 
85
- The issue comment and `docs/solutions/` commit are the only retro outputs. Never write a
86
- handoff, phase artifact, local feedback log, or completion label.
97
+ The Dispatch message and the `docs/solutions/` commit are the only retro outputs. Never write a
98
+ handoff, phase artifact, local feedback log, or completion label; `.legion/` was deleted before
99
+ retro and nothing recreates it. Report completion with `legion handoff complete` alone (its
100
+ summary: two sentences for the architect) — no `legion handoff write`.
87
101
 
88
102
  ## Completion check
89
103
 
90
104
  Before returning, verify all of the following:
91
105
 
92
- - The reviewer cleanup commit remains below the retro documentation commit.
93
- - The learning documents and issue comment both exist.
106
+ - The reviewer cleanup commit remains below the retro documentation commit, and the reviewer's
107
+ approval of that cleanup head stands: the merger accepts the approved head plus this commit.
108
+ - The learning documents and the Dispatch message both exist.
94
109
  - No `.legion` file was created or modified by retro.
95
110
  - The fresh-eyes analysis was considered alongside the implementer's context.
96
111
  - The merger remains a subsequent step, not work performed by retro.
@@ -89,7 +89,7 @@ Concurrent issues have disjoint workspaces; only the currently active phase muta
89
89
  one. After you complete and go idle, treat `$LEGION_WORKSPACE` as read-only: you are kept
90
90
  alive to answer questions, not to keep editing. Do not create new commits, run
91
91
  `jj -R "$LEGION_WORKSPACE" new`, or touch tracked files once your own handoff is committed
92
- and pushed — a code change belongs to whichever phase is active now.
92
+ (and, for the implementer, pushed) — a code change belongs to whichever phase is active now.
93
93
 
94
94
  On every start, and especially after revival or re-creation, read the issue and then the
95
95
  committed predecessor handoffs in lifecycle order from `$LEGION_WORKSPACE/.legion/`:
@@ -120,10 +120,21 @@ transcript, the committed file wins: it is the copy that survived.
120
120
  - **Before pushing, check ancestry:** `jj -R "$LEGION_WORKSPACE" log -r 'ancestors(@, 5)'`
121
121
  — verify only your issue's commits are in the chain, not unrelated work.
122
122
 
123
- **Shared operation safety:** Never run `jj op restore` in a Legion workspace. It rewrites
124
- the shared operation log. If a mistake reaches that point, stop and send the owning
125
- architect the `jj -R "$LEGION_WORKSPACE" log` evidence; recover only through the approved,
126
- path-scoped workflow.
123
+ **Shared operation safety:** Every Legion issue workspace is a `jj workspace` of one shared
124
+ clone, so they all share one operation log: `jj undo`, `jj abandon`, and
125
+ `jj op restore|revert|abandon|undo` rewrite it for every tree at once (on 2026-09-12 one
126
+ worker's `jj undo` rewrote nine of another tree's commits). The extension refuses them in every
127
+ phase-worker pane before they run — a `bash` command in any position of a pipeline or `&&`
128
+ chain, with or without `-R`, judged on the whole argument list; `eval` code; and a `hub`
129
+ process start — from your own tool calls and from any `task` subagent you spawn (it runs in
130
+ your pane, against the same log), and a `bash` command whose quoted text merely mentions `jj`
131
+ with one of those words (a heredoc, an echo, a commit message) is refused too: write such text
132
+ with the `write` tool or say "operation-log rollback" instead. `jj restore <paths>`,
133
+ `jj op log`, and `jj op show` stay allowed. Recover forward only: a new commit
134
+ (`jj -R "$LEGION_WORKSPACE" new`) or `jj -R "$LEGION_WORKSPACE" restore <paths>` of files.
135
+ Anything else, stop and send the owning architect the `jj -R "$LEGION_WORKSPACE" log`
136
+ evidence; the architect decides, and an operator performs any operation-log restore with every
137
+ other tree paused.
127
138
 
128
139
  ## Phase work
129
140
 
@@ -138,9 +149,22 @@ Commit attribution is automatic: the extension exports a `JJ_CONFIG` overlay whe
138
149
  session starts, so every jj commit you make carries an `Omp-Session: <this-session-id>`
139
150
  trailer with no action from you. Do not add attribution trailers by hand.
140
151
 
141
- The jj configuration already supplies your phase's plus-addressed author and committer
142
- identity. Do not override Git identity configuration. Your session receives the credential
143
- capability it needs; invoke GitHub through the credential helper:
152
+ Your pane's environment already supplies your phase's author and committer identity
153
+ (`JJ_USER`/`JJ_EMAIL` and the Git author/committer variables, set by the daemon when it opened
154
+ the pane; the daemon also re-authors the workspace's working copy for your role at each
155
+ assignment, since `jj split`/`jj describe` keep its author). Never set or override
156
+ `user.name`/`user.email` in any jj or Git scope — not `jj config set`, not `--config`, not
157
+ `git config`: `--config` outranks the pane environment and would put the wrong App back on your
158
+ commits, and the repository-scoped jj config is one file shared by every issue workspace of the
159
+ clone. Before a push, check
160
+ `jj -R "$LEGION_WORKSPACE" log -r 'main@origin..@' -T 'author.email() ++ " | " ++ committer.email() ++ " " ++ description.first_line() ++ "\n"'`
161
+ shows your role's App in both columns **on every commit you made** — not on the whole list:
162
+ earlier phases' commits are legitimately authored by their own role's App, and a conflict-forced
163
+ rebase legitimately sets the committer of every rebased commit, other roles' included, to the
164
+ rebaser. A wrong identity on your own commit is a pane-environment problem to report to the
165
+ architect, not something to pin (`docs/solutions/legion/shared-main-repo-hazards-for-concurrent-issue-workspaces.md`,
166
+ Hazard 1). Your session receives the credential capability it needs; invoke GitHub through the
167
+ credential helper:
144
168
 
145
169
  ```bash
146
170
  legion gh -- <gh args…>
@@ -152,22 +176,29 @@ that execs `legion gh -- "$@"`, so `gh …` and `legion gh -- …` are the same
152
176
  redeems a fresh token from your session's grant — identity is supplied per call, never stored.
153
177
  Never run `gh auth login` or `gh auth setup-git`; there is no login state to create. The shim
154
178
  refuses `pr merge` (and a raw `gh api …/merge`): no worker role merges a pull request — the merge
155
- queue does, under its own authority. The credential reaches `legion` through the file
156
- `$LEGION_GRANT_FILE` names, written before each of your bash commands by the extension; never
157
- `cat`, `echo`, copy, or `export` it `legion credential`, `legion gh`, `jj git push`, and
158
- `legion handoff complete` read it themselves. The file is the pane's, not the command's: a `task`
159
- subagent, an `eval` subprocess, or a background job in your pane reads the grant your last bash
160
- command minted, so its `legion gh` or `jj git push` succeeds only within 60 seconds of that call
161
- and 403s afterwards a timing artifact, not a broken credential; run credentialed commands from
162
- your own bash calls.
179
+ queue does, under its own authority. It also refuses every GitHub-issue write — the `issue`
180
+ subcommand's `comment`, `create`, `edit`, `close`, `reopen`, `delete`, `pin`, `unpin`, `transfer`,
181
+ `lock`, `unlock`, and `develop`, and any raw `gh api` call to an `/issues` path whose method is not
182
+ GET (an explicit `-X`, or the POST that `-f`/`-F`/`--input` imply; pull-request conversation
183
+ comments live on that path too, so edit them with `gh pr comment`) printing
184
+ `Legion issues live on Dispatch; use dispatch_message or dispatch_comment on <your LEGION_ISSUE>`:
185
+ Legion never reads or writes a GitHub issue (LEGION-78). `pr comment`, `pr review`,
186
+ `api …/pulls/…`, `api graphql`, and issue reads are unaffected. The credential reaches `legion`
187
+ through the file `$LEGION_GRANT_FILE` names, written before each of your bash commands by the
188
+ extension; never `cat`, `echo`, copy, or `export` it — `legion credential`, `legion gh`,
189
+ `jj git push`, and `legion handoff complete` read it themselves. The file is the pane's, not the
190
+ command's: a `task` subagent, an `eval` subprocess, or a background job in your pane reads the
191
+ grant your last bash command minted, so its `legion gh` or `jj git push` succeeds only within 60
192
+ seconds of that call and 403s afterwards — a timing artifact, not a broken credential; run
193
+ credentialed commands from your own bash calls.
163
194
 
164
195
  ## GitHub PR comment attribution
165
196
 
166
197
  Append this exact structured footer to **every** pull-request comment and review that this
167
198
  phase posts on GitHub. It preserves session provenance on the artifact itself so work stays
168
199
  attributable to the session that produced it. Dispatch comments carry session provenance
169
- natively through their own `actor`/`origin` fields; this footer is only for GitHub PR
170
- artifacts:
200
+ natively through their own `actor`/`origin` fields; this footer is for GitHub PR artifacts and
201
+ for the retro's Dispatch message (`skills/legion-retro`):
171
202
 
172
203
  ```html
173
204
  <!-- legion: {"session":"<session-id>","phase":"<phase>"} -->
@@ -220,11 +251,14 @@ PR opens, and every later phase keeps it current rather than replacing it:
220
251
  ```
221
252
  ## Verification
222
253
 
223
- **CI:** `pr-checks-result` run <run-id> — success at <head-sha>.
254
+ **CI:** `Tests` run <run-id> — jobs lint, pr-title, typecheck, test all success at <head-sha>.
224
255
 
225
256
  **Threads:** <n> resolved, 0 unresolved. Each disposed individually, never in bulk:
226
257
  - Thread <id>: fixed in <commit-sha> — <one line>.
227
258
  - Thread <id>: not a defect — <reason>.
259
+ `legion threads resolve --pr <n> --repo <owner>/<repo>` at <head-sha>:
260
+ resolved <thread URL>
261
+ left open <thread URL> — newest reply by <login> is not an acceptance
228
262
 
229
263
  **Thermo:** thermonuclear-deep-review + thermonuclear-code-quality run once at <head-sha>:
230
264
  <verdict>. (omitted entirely on a docs-only PR — no thermo pass runs)
@@ -238,8 +272,24 @@ Negative control: <deliberately broken input> → <refusal or failure observed>.
238
272
  ```
239
273
 
240
274
  - **Threads are dispositioned individually, never resolved in bulk.** Every open review
241
- thread gets its own line naming the fixing commit or the reason it isn't a defect before
242
- it is marked resolved.
275
+ thread gets its own line naming the fixing commit or the reason it isn't a defect. The
276
+ reviewer answers each thread it opened with exactly one of `Accepted: fixed in <commit> — <one line>`,
277
+ `Accepted: not a defect — <reason>`, or `Still open: <what remains>`; nothing else is an
278
+ acceptance, and nobody replies after an `Accepted:` (any later reply that is not itself an
279
+ `Accepted:` — the opener's own follow-up included — leaves the thread open, since the command
280
+ reads only the newest comment). The review App can reply on a thread but can neither resolve it
281
+ nor push — GitHub grants both only to the pull request's author or an account with write (push)
282
+ access to the repository, and the review App is neither by design
283
+ (`packages/daemon/src/daemon/AGENTS.md`, GitHub Apps) — so the
284
+ **implementer** runs `legion threads resolve --pr <number> --repo <owner>/<repo>` before every
285
+ push that answers a review (the corrective push and the final `.legion/` deletion push) and
286
+ pastes its output into the `Threads` section. The command resolves each unresolved thread
287
+ whose newest comment is the opener's own `Accepted:` reply, one `resolveReviewThread` per
288
+ thread, prints `resolved <url>` or `left open <url> — newest reply by <login> is not an acceptance`,
289
+ and exits 1 naming the thread's URL and GitHub's message when GitHub refuses one; report that
290
+ exit to the architect, which opens an action ask for a human to resolve the thread by hand —
291
+ never skip it silently. The merger runs the same command once more before publishing READY
292
+ and does not publish while any `left open` line remains.
243
293
  - **Correctness fixes land in this PR; cleanup is one named fast-follow.** A finding that
244
294
  changes behavior, hides an error, or breaks a gate is fixed here — never deferred.
245
295
  Findings about naming, duplication, or wording are batched into the single `Fast-follow`
@@ -249,7 +299,13 @@ Negative control: <deliberately broken input> → <refusal or failure observed>.
249
299
  The implementer rebases the issue branch only when GitHub reports it `CONFLICTING` or the
250
300
  controller asks because of a conflict — never to pick up `main` or to refresh CI. A single
251
301
  failed CI job is re-run on its own with `legion gh -- run rerun <run-id> --failed`, never by
252
- pushing a new commit.
302
+ pushing a new commit. A conflict-forced rebase that leaves the branch's diff unchanged is a
303
+ confirmation, not a new round (see *The unchanged-diff check* below). Before rebasing, record
304
+ the fingerprint at the current tip; after pushing the rebased branch, record it at the new
305
+ tip; post one PR comment (Legion footer):
306
+ `rebase <old-tip-sha> → <new-tip-sha>; fingerprint <before> → <after>; unchanged|changed`.
307
+ Rebase the whole chain — `jj -R "$LEGION_WORKSPACE" rebase -s 'roots(main@origin..@)' -d main@origin` —
308
+ so the tester's and reviewer's commits move with yours.
253
309
  - **No deferrals.** Sami, 2026-09-11, verbatim: "My rule is no deferrals." The `Fast-follow:`
254
310
  field names naming, duplication, or wording cleanup only; anything that changes behaviour,
255
311
  hides an error, or breaks a gate lands in this PR.
@@ -275,25 +331,56 @@ Negative control: <deliberately broken input> → <refusal or failure observed>.
275
331
  whole staging gate never ran. Environment or
276
332
  secret-scrub evidence (e.g. "`LEGION_*`/`DISPATCH_*`/`ENVOY_*` unset") is recorded once, in
277
333
  `.legion/test.json`, and only when the issue's acceptance criteria call for it — never
278
- re-pasted into the PR body each round.
334
+ re-pasted into the PR body each round. After a conflict-forced rebase, compute the
335
+ fingerprint at the head your `E2E` line names and at the new head. Equal: re-run only the
336
+ bare gates — the repository's CI green at the new head and its smoke check — and change the
337
+ `E2E` line's head to the new SHA with
338
+ `rebase re-check <old-sha> → <new-sha>: fingerprint unchanged, bare gates only`; the
339
+ real-surface verification is not repeated. Different: a full test round.
279
340
  - The reviewer verifies the `CI`, `Threads`, and `E2E` facts against GitHub directly —
280
341
  never from a handoff — then runs `task(agent="thermonuclear-deep-review")` and
281
342
  `task(agent="thermonuclear-code-quality")` once at that head and records the verdict.
282
343
  Skip the `Thermo` line entirely on a docs-only PR. Submit **one review per round** —
283
344
  `REQUEST_CHANGES` when any correctness finding stands, otherwise `COMMENT` while the head
284
- still carries `.legion/`; `APPROVE` only for the head that differs from the reviewed one by
285
- the `.legion/` deletion alone, named by SHA — carrying every inline comment in that single
345
+ still carries `.legion/`; `APPROVE` only for a head that carries no `.legion/` the head
346
+ that differs from the reviewed one by the `.legion/` deletion alone, or, after a
347
+ conflict-forced rebase, the new head whose fingerprint equals the approved head's — always
348
+ named by SHA — carrying every inline comment in that single
286
349
  call: `legion gh -- api --method POST repos/{owner}/{repo}/pulls/{number}/reviews --input body.json`
287
350
  with `commit_id`, `event` (`REQUEST_CHANGES`, `COMMENT`, or `APPROVE`), `body` (with the
288
351
  Legion footer), and a `comments[]` array of `{path, line, side, body}`, one entry per
289
352
  finding — never one `pr review` call per finding (each submission fires a `pr-review` wake).
290
353
  Then return the issue to the architect; when clean, have the architect send the implementer
291
354
  back to push the `.legion/` deletion (the review App cannot push), then review **that** head
292
- and approve it by name.
355
+ and approve it by name. After a conflict-forced rebase, compute the fingerprint at the
356
+ `commit_id` of your last submitted review and at the new head. Equal and that review was
357
+ `APPROVE`: submit one more `APPROVE` naming the new head by SHA, its body naming both SHAs
358
+ and the fingerprint — a confirmation, not a round; no thermo pass, no thread pass. Equal and
359
+ that review was `COMMENT` or `REQUEST_CHANGES`: continue that round against the new head;
360
+ nothing restarts. Different: a new round — thermo again, one review.
361
+ When you re-review after a corrective push, answer every thread you opened in one of the
362
+ three forms above — `Accepted:` is the only reply the implementer's `legion threads resolve`
363
+ acts on — and approve only once every thread you opened carries your `Accepted:` reply and the
364
+ implementer's run has resolved it (verify `isResolved: true` with `gh api graphql`, never from
365
+ the PR body).
293
366
  - Once a base is frozen for others to stack on, never rewrite it — fixes land as new
294
367
  commits on top, and the `Chain` line records what is frozen.
295
- - The merger confirms the approved head still equals the current head, then publishes
296
- `READY #<n> at <sha>` plus the PR body's gate facts to the merge queue's role
368
+ - **Retro's commit does not void the reviewer's approval.** After the reviewer approves the
369
+ cleaned head, retro commits its learnings under `docs/solutions/` on top of it; that commit
370
+ stays, the approval stands, and the tree goes to the merger — never back to the tester or
371
+ reviewer. Anything else above the approved head does void it, and the merger tells the
372
+ architect the head must return to review instead of publishing. A conflict-forced rebase
373
+ after retro moves those documents with the branch; retro never re-runs.
374
+ - The merger runs `legion threads resolve --pr <n> --repo <owner>/<repo>` (it acts as the same
375
+ code-writing App as the implementer; resolving a thread changes no commit, so this run never
376
+ invalidates the approval), does not publish while any `left open` line remains or the command
377
+ exits 1 (report the thread to the architect instead), then
378
+ proves that rule with two commands and publishes. First
379
+ `cd -- "$LEGION_WORKSPACE" && jj -R "$LEGION_WORKSPACE" git fetch && jj -R "$LEGION_WORKSPACE" diff --from <approved-sha> --to <tip-sha> --summary`,
380
+ whose output is quoted in READY (an empty output is quoted as
381
+ `no file changes above the approved head`); then the same with `'~docs/solutions'` appended,
382
+ which must print nothing. Then it publishes `READY #<n> at <tip-sha>` naming the approved
383
+ head, the tip, and that summary, plus the PR body's gate facts, to the merge queue's role
297
384
  (`notifications.role.pr-queue`) with `envoy_publish`. The merger never merges; the queue
298
385
  merges under its own authority.
299
386
  - **After the queue merges, the implementer verifies in production.** Sami, 2026-09-13,
@@ -306,6 +393,44 @@ Negative control: <deliberately broken input> → <refusal or failure observed>.
306
393
  failed at 00:12Z on a resource staging never runs. If the slot fails on the change, the
307
394
  implementer owns the fix and the next slot.
308
395
 
396
+ ## The unchanged-diff check
397
+
398
+ The fingerprint every role compares after a conflict-forced rebase (every flag and the fileset
399
+ verified on jj 0.45.1):
400
+
401
+ ```bash
402
+ cd -- "$LEGION_WORKSPACE" && jj -R "$LEGION_WORKSPACE" git fetch && \
403
+ jj -R "$LEGION_WORKSPACE" diff --from "fork_point(main@origin | <head-sha>)" --to <head-sha> \
404
+ --git --context 0 '~(.legion | docs/solutions)' \
405
+ | sed -e '/^@@/d' -e '/^index /d' | sha256sum
406
+ ```
407
+
408
+ - `<head-sha>` is a full commit SHA; a jj commit id is the git SHA GitHub shows.
409
+ - `fork_point(main@origin | <head-sha>)` is the base the branch was cut from *at that head*:
410
+ the old base for the pre-rebase head, the new base for the rebased one, so one command
411
+ serves both sides. On a stacked PR substitute its base branch for `main`
412
+ (`legion gh -- pr view <n> --json baseRefName`).
413
+ - A head the rebase hid is still addressable by its SHA in the shared workspace. A SHA the
414
+ workspace cannot resolve (`jj -R "$LEGION_WORKSPACE" log -r <sha>` errors) counts as a
415
+ changed diff — never as unchanged.
416
+ - `--context 0` drops context lines; the `sed` drops `@@` hunk headers (line positions move
417
+ on a rebase) and `index` lines (blob ids move when the base's copy of a file changed). What
418
+ is left is exactly the added and removed lines per file.
419
+ - The single fileset `'~(.legion | docs/solutions)'` leaves out the handoff ledger and retro's
420
+ learnings: process artifacts the rules above already exempt from re-review, which change
421
+ between one role's verified head and the next without changing the product. This is what lets
422
+ each role compare against *its own* last verified head instead of trusting another role's
423
+ numbers. It must be one expression: jj unions positional filesets, so two separate
424
+ `'~.legion' '~docs/solutions'` arguments select every file and exclude nothing. Once
425
+ `.legion/` is gone, jj warns `No matching entries for paths: .legion` on stderr; the hash is
426
+ unaffected.
427
+
428
+ Where each role gets its two heads: the implementer — the tip before and after its own rebase;
429
+ the tester — the head its `E2E` line names and the new head; the reviewer — the `commit_id` of
430
+ its last submitted review (`legion gh -- api repos/{owner}/{repo}/pulls/{n}/reviews --jq '.[] | {commit_id, state, user: .user.login}'`)
431
+ and the new head; the merger never computes a fingerprint — it uses the `--summary` check
432
+ above.
433
+
309
434
  ## Completion gate: handoff write, verification, and persistence
310
435
 
311
436
  Write the phase-specific handoff:
@@ -328,9 +453,12 @@ cd -- "$LEGION_WORKSPACE" && \
328
453
  jj -R "$LEGION_WORKSPACE" split -m "<phase>: record handoff" .legion/<phase>.json
329
454
  ```
330
455
 
331
- If the issue bookmark exists locally, advance it and push it with the provisioned
332
- credential helper. `--bookmark` also publishes the locally provisioned bookmark on its
333
- first push a bookmark not yet tracking a remote one is tracked automatically:
456
+ **Only the implementer pushes the issue branch.** It acts as the code-writing App
457
+ (`legion-implementer[bot]`, `appRoleForLegionRole` in `packages/daemon/src/daemon/github-apps.ts`),
458
+ the one App with `contents` permission (the merger acts as the same App but pushes nothing: it
459
+ verifies and publishes READY). If you are the implementer, advance the issue bookmark and push it
460
+ with the provisioned credential helper. `--bookmark` also publishes the locally provisioned
461
+ bookmark on its first push — a bookmark not yet tracking a remote one is tracked automatically:
334
462
 
335
463
  ```bash
336
464
  cd -- "$LEGION_WORKSPACE" && \
@@ -338,10 +466,23 @@ cd -- "$LEGION_WORKSPACE" && \
338
466
  jj -R "$LEGION_WORKSPACE" git push --bookmark legion/<KEY>
339
467
  ```
340
468
 
341
- Do not report phase completion until the write, existence check, and handoff commit
342
- succeed; when an issue branch exists, its push is also required. This is the committed
343
- copy the next phase reads after revival. It is removed once, at the end of a clean review: the
344
- implementer pushes that deletion at the reviewer's direction. No other phase removes it.
469
+ Every other role planner, tester, reviewer, architects acts as the review App
470
+ (`legion-reviewer[bot]`), which cannot push: the `split` above is your last step, and the commit
471
+ rides the implementer's next push (the corrective push after a review, or the final `.legion/`
472
+ deletion). A push from one of those roles is refused over git it reads
473
+ `remote: Repository not found.`; the REST API's form of the same refusal is
474
+ `Resource not accessible by integration` — and that refusal is expected, not a failure to report
475
+ or retry.
476
+
477
+ Do not report phase completion until the write, existence check, and handoff commit succeed —
478
+ and, for the implementer, until the push has too. This is the committed copy the next phase
479
+ reads after revival. It is removed once, at the end of a clean review: the implementer pushes
480
+ that deletion at the reviewer's direction. No other phase removes it — and once it is gone
481
+ (`jj -R "$LEGION_WORKSPACE" file list -r @- .legion` prints nothing on stdout; jj warns on
482
+ stderr), this gate no longer applies: a later rebase, bare-gate re-check, confirmation, or retro
483
+ writes no `.legion/<phase>.json`, commits no handoff, and reports with `legion handoff complete`
484
+ alone (below). Recreating `.legion/` after its deletion changes the approved head and restarts
485
+ the review loop this rule exists to end.
345
486
 
346
487
  ## Completion: report to the architect, then stay
347
488