opencode-plugin-flow 6.7.0 → 6.8.0

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.
package/dist/index.js CHANGED
@@ -6,153 +6,103 @@ description: Drive a Flow goal from planning through implementation, validation,
6
6
 
7
7
  # Flow
8
8
 
9
- Flow is a small state ledger around coding work. An active Flow session is
10
- authoritative for its goal until completed, deferred, or abandoned closure.
11
- Do not silently fall back to ordinary non-Flow coding. The root manager owns
12
- the session, integration, validation, review dispatch, reset, closure, and every
13
- manager-owned lifecycle mutation. Bounded \`flow-worker\`
14
- instances may contribute disjoint work inside the active feature. The reserved
15
- \`flow-reviewer\` independently reviews and submits its own result through
16
- \`flow_feature_complete\`; it cannot edit the workspace or make any other
17
- lifecycle mutation. Do not dispatch generic or general-purpose agents for
18
- active Flow planning, implementation, evidence gathering, or review. Use only
19
- the root manager and the two reserved Flow roles.
9
+ Flow's active goal is authoritative until completed, deferred, or abandoned;
10
+ never silently fall back.
20
11
 
21
12
  ## Route from status
22
13
 
23
14
  1. Call \`flow_status { request: { view: "compact" } }\` first. Trust its
24
- projection over conversation memory. Treat \`nextAction\` as the durable
25
- default workflow direction, not permission to exceed or a reason to discard
26
- existing user authority.
27
- If compact status contains \`archiveRetry\`, the close was already accepted:
28
- call \`flow_session_close\` once with that projected request byte-for-byte,
29
- then refresh compact status. Stop if archive publication remains
30
- unconfirmed; otherwise continue from the refreshed projection. This exact
31
- cleanup grants no new work, so it precedes goal alignment.
32
- Before any other manager-owned lifecycle mutation, align the compact-projected
33
- goal with the current user request. Continuation and compatible narrowing
34
- proceed inside the approved goal. Compatible narrowing changes method or
35
- emphasis only; it must not add, drop, reorder, or weaken an approved
36
- requirement or feature outcome. For materially new or expanded work,
37
- perform no mutation. Say the new request has not started, and offer to
38
- continue the active goal, defer it, or abandon it. A completed-but-unclosed
39
- session must close as completed before the new request proceeds. This
40
- comparison is conversational only: create no queue, classifier, or new
41
- state. A durable \`nextAction\` can still be rejected after status by an
42
- environment-sensitive guard; refresh compact status and handle the exact
43
- rejection instead of forcing a stale action.
44
- 2. If the user asked only for a plan and an approved same-goal session already
45
- exists, load \`flow_status { request: { view: "detail" } }\` once. Report the
46
- immutable active plan and current progress, then stop. Do not call
47
- \`flow-plan\` or \`flow-run\`.
48
- 3. If there is no session or the plan is still a draft, call
49
- \`flow_guidance { id: "flow-plan" }\` and follow that contract. Stop after
50
- planning when the user asked for a plan only.
51
- 4. If an approved feature is ready, running, or blocked, call
52
- \`flow_guidance { id: "flow-run" }\` and follow that contract for exactly that
53
- feature, including its retry and checkpoint routing.
54
- 5. After applying \`flow-run\` to one feature, read compact status again. Reuse
55
- its one detail projection for any blocked checkpoint. Under \`/flow-auto\`,
56
- a passing feature and a compact status of \`ready\` or \`completed\` are loop
57
- states, not handoff states: immediately apply \`flow-run\` to the next ready
58
- feature, or close the completed session with one \`flow_session_close\`
59
- request. Do not return “ready for the next feature,” offer \`/flow-run\`, or
60
- wait for another user turn while an authorized next action is runnable. For
61
- a blocked outcome, follow the loaded \`flow-run\` retry and checkpoint
62
- contract.
15
+ \`nextAction\` as the durable default, not added authority.
16
+ If the top-level response is an error, report its exact summary and recovery;
17
+ when delivery is present, handle its bounded map under **Recovery**. Stop
18
+ without another mutation.
19
+ For \`archiveRetry\`, call the projected \`flow_session_close\` request
20
+ byte-for-byte and handle its delivery under **Recovery**. Refresh compact status;
21
+ stop if publication remains unconfirmed, otherwise continue from the
22
+ refreshed projection. This cleanup precedes goal alignment and grants no work.
23
+ Before another manager mutation, align the compact-projected goal with the
24
+ request. Continue only for the same goal or a narrowing that preserves every
25
+ outcome. Close a completed session as completed before new work. If the user
26
+ explicitly chooses deferred or abandoned closure for a non-completed session,
27
+ call \`flow_session_close\` with compact session id/revision, fresh operation
28
+ id, that kind, and optional summary; handle **Recovery**, follow a projected
29
+ exact \`archiveRetry\`, and stop. Otherwise new or expanded work makes no
30
+ mutation: say it has not started and offer continue, defer, or abandon. Keep
31
+ alignment conversational. On revision conflict, refresh compact status and
32
+ retry only after confirming the same session and goal and that status still
33
+ permits the selected closure kind; never close a replacement.
34
+ 2. For a plan-only request with an approved same-goal session, read detail once,
35
+ report the immutable plan and progress, and stop.
36
+ 3. With no session or a draft, load \`flow_guidance { id: "flow-plan" }\`; stop
37
+ after planning if implementation was not authorized.
38
+ 4. For an approved ready, running, or blocked feature, load
39
+ \`flow_guidance { id: "flow-run" }\` and follow it for exactly that feature.
40
+ 5. After \`flow-run\`, reload compact status. Route blocked outcomes through the
41
+ loaded retry and checkpoint contract; otherwise continue **End-to-end loop**.
63
42
 
64
43
  ## End-to-end loop
65
44
 
66
- Once implementation is authorized, \`/flow-auto\` treats compact \`ready\` and
67
- \`completed\` as mechanical loop states. A host-triggered continuation begins
68
- from a provisional compact baseline: it may route only after the initiating
69
- turn creates a Flow session from idle or advances that same Flow session.
70
- Command entry alone does not authorize an unchanged pre-existing ready session,
71
- and a replacement session fails closed. Conversational \`flow_plan_approve\` and
72
- blocked or ready \`await-user-direction\` replies remain part of the same Flow
73
- interaction, but auto-routing resumes only after the reply advances that same
74
- session to a mechanical state.
75
-
76
- 1. For \`ready\`, load or reuse \`flow-run\` and run the projected feature.
77
- 2. After every recorded result, reload compact status and route again.
78
- 3. For \`completed\`, close and finish any exact \`archiveRetry\`.
79
-
80
- Intermediate progress and \`nextAction: "flow_run_start"\` are not terminal.
81
- Return only after closure or a required checkpoint. Direct \`/flow-run\` still
82
- stops after one feature.
83
-
84
- Within existing implementation authority, continue after approval and every
85
- passing feature outcome without asking again. For a blocked outcome, follow the
86
- loaded \`flow-run\` retry and checkpoint contract. The session remains
87
- authoritative while blocked. Never implicitly select a feature whose latest
88
- relevant reviewed outcome remains failed. Untouched dependency-independent
89
- features may continue; if only retry-required candidates remain, wait at
90
- \`await-user-direction\` for the exact choice. If status is blocked, carry the
91
- authorized retry or independent choice atomically as \`nextFeatureId\` on
92
- \`flow_feature_reset\`. If status is ready, the failed run is already superseded:
93
- load detail once and use \`flow_run_start\` with the explicitly authorized
94
- retry's exact \`featureId\`. Never reset from ready status, and add no hold or
95
- retry ledger.
96
-
97
- Pause only for a convergence checkpoint, a material product or scope choice,
98
- missing authority for an external Git or release action,
99
- a hard operational failure, or the user's explicit choice of deferred or
100
- abandoned closure. Only the user may choose either non-completed kind.
101
-
102
- Core contracts are bundled in the plugin; load them through \`flow_guidance\` and
103
- do not depend on native skill discovery. If a required Flow tool is unavailable,
104
- report that the plugin is not fully loaded instead of simulating state changes.
45
+ \`/flow-auto\` loops \`ready\`/\`completed\` only after a same-host non-replayed
46
+ \`flow_plan_save\` owns idle session creation, or an active baseline advances in
47
+ that session. The latter temporal gate admits a pending reviewer result;
48
+ unchanged ready or replacement fails closed;
49
+ conversational \`flow_plan_approve\` and blocked or ready \`await-user-direction\`
50
+ resume only when the reply's same-host accepted mutation advances it.
51
+
52
+ For \`ready\`, apply \`flow-run\`; after every recorded result reload compact. For \`completed\`,
53
+ close and handle **Recovery** plus exact \`archiveRetry\`.
54
+ \`flow_run_start\` and intermediate progress are not terminal: never return “ready
55
+ for the next feature” or wait while an action is authorized. Return only after closure
56
+ or a checkpoint. Direct \`/flow-run\` stops after one feature.
57
+
58
+ For blocked work, follow \`flow-run\` routing. Never implicitly select a feature
59
+ whose latest relevant reviewed outcome remains failed.
60
+ Untouched independent features may continue; if only retries remain, wait at
61
+ \`await-user-direction\`. Otherwise pause only for a material
62
+ choice, missing Git/release authority, hard failure, or user-chosen closure.
63
+ If a Flow tool is absent, report an incomplete plugin load; never simulate state.
105
64
 
106
65
  ## Invariants
107
66
 
108
- - Approved plans do not change. If implementation requires material scope
109
- outside the plan, stop editing. Finish the approved plan or have the user
110
- explicitly choose deferred or abandoned closure before starting a new plan;
111
- do not replan in place.
112
- - Only one durable feature run is active at a time. Conversation-local worker
113
- waves do not create additional runs or Flow state.
114
- - Work stays inside the active feature and preserves unrelated user changes.
115
- - A passing feature needs successful current-source validation and one result
116
- submitted directly by the assigned independent reviewer. The final feature
117
- uses broad validation and a final review; it does not add a second review
118
- pass.
119
- - A failed review is recorded honestly. Any retry is a fresh run with full
120
- validation and review.
121
- - Use runtime revisions and operation ids for ordering and idempotency. Supply
122
- only fields requested by the current tool schema.
123
- - Do not create reports or sidecars outside \`.flow/**\` unless the user asks for
124
- a durable artifact. Prefer one readable Markdown file; JSON is opt-in.
125
- - Do not stage, commit, push, publish, or mutate releases unless the user
126
- explicitly asks for that Git or release action.
127
-
128
- ## Blocked handoff
129
-
130
- A blocked handoff must be self-contained and label the result overall
131
- incomplete. Use the one detail projection for a failed-review block. Report the
132
- goal and progress; blocked feature, attempt, failure count, findings; completed
133
- and untouched features; validation and artifact evidence; Git and release
134
- mutation status; and whether the newest request started and matched the active
135
- goal.
67
+ - Plans are immutable. Out-of-plan work stops; finish or explicitly close, never
68
+ replan in place.
69
+ - One durable feature run exists at a time; worker waves add no Flow state.
70
+ - Preserve unrelated work and stay inside the active feature.
71
+ - Passing requires current-source validation and one independent review; the
72
+ final feature uses broad validation, not another review.
73
+ - Record failures honestly; retries are fresh full runs.
74
+ - Use runtime revisions, fresh operation ids, and current-schema fields.
75
+ - Do not create reports/sidecars, Git changes, or release mutations unless
76
+ requested; reports default to Markdown and JSON is opt-in.
136
77
 
137
78
  ## Recovery
138
79
 
139
- On confusion or interruption, read compact status and route with the rules
140
- above; load \`flow-run\` for an active or blocked feature and apply its exact
141
- review-recovery path. Use execution status for active work and reviewer status
142
- for a returned assignment id. Never invent or submit a verdict, or infer
143
- completion, retry count, or closure from prose.
144
-
145
- For a newly completed session, close with one request containing the
146
- status-projected session id, a fresh operation id, current revision, closure
147
- kind, and optional summary. Repeating that exact request converges; there is no
148
- separate retry mode.
149
-
150
- After a durably accepted close, build the final handoff from
151
- \`workflowData.delivery\`. For each feature, report its attempt count, latest
152
- outcome, and terminal findings. Label its artifact groups Flow-reported
153
- artifacts from latest attempts and Flow-reported artifacts from superseded
154
- attempts only. Never describe them as an exact Git delta. Do not create a
155
- report unless the user asks for one.
80
+ On interruption, read compact status; load \`flow-run\` for an active or blocked feature
81
+ and use execution or reviewer status, never prose, for lifecycle truth.
82
+
83
+ Summaries keep plan/source IDs \`verified\` or \`incomplete\`. A prior finding is
84
+ terminally \`fixed\` only when review passes with current evidence. A failed
85
+ verdict carries every prior ID forward: report a proven repair as \`terminal
86
+ fixed pending pass\` with concise evidence, an
87
+ unproven repair as unverified-fixed, a recurrence as \`recurring\`, or a confirmed
88
+ nonblocker as \`residual\`. Blockers stay terminal.
89
+
90
+ For an accepted close, map only \`workflowData.delivery\`
91
+ \`outcomeSummary\`/\`terminalFindings\`.
92
+ Requirements are proven \`verified\`, otherwise \`incomplete\` or explicit
93
+ \`deferred\`. Apply the finding rules above, using \`incomplete\` for an unproven
94
+ terminal claim; \`abandoned\` stays the kind. Missing IDs are unavailable: never
95
+ invent or read detail solely for closure. Without delivery, report exact
96
+ recovery and no map.
97
+
98
+ Unresolved blockers forbid completed closure. Fresh close: projected session
99
+ id/revision, fresh operation id, kind, optional summary. Replay byte-for-byte
100
+ only the \`archiveRetry\` of a durably accepted close. Rejected revision conflict:
101
+ refresh compact, confirm the same session/goal, then build a fresh request.
102
+ From delivery report each feature's attempt count, latest outcome, terminal
103
+ findings, and Flow-reported latest/superseded artifacts. Report external prerequisites only
104
+ from terminal text; otherwise mark them unavailable. Artifacts are not an exact
105
+ Git delta. Create no other ledger or report.
156
106
  `;
157
107
 
158
108
  // skills/flow-plan/SKILL.md
@@ -163,37 +113,46 @@ description: Create, revise, or approve a concise Flow plan before implementatio
163
113
 
164
114
  # Flow Plan
165
115
 
166
- Plan only after reading the repository facts that determine the work. A useful
167
- plan is short enough to scan and concrete enough that another agent can execute
168
- it without rediscovering the goal.
116
+ Inspect determining repository facts. Keep the plan scannable and executable
117
+ without rediscovering the goal.
169
118
 
170
119
  ## Start
171
120
 
172
121
  - Call \`flow_status { request: { view: "compact" } }\` first.
173
- - If compact status contains \`archiveRetry\`, finish that already-accepted close
174
- by calling \`flow_session_close\` once with the projected request byte-for-byte,
175
- then refresh compact status. Stop without saving a plan if archive publication
176
- remains unconfirmed; otherwise continue from the refreshed projection. This
177
- exact cleanup grants no new work, so it precedes goal alignment. Apply the
178
- same rule if closing a completed session later returns archive-pending.
179
- - Before any other manager-owned lifecycle mutation, align the compact-projected
180
- goal with the current direct \`/flow-plan\` request. Continue only for the same
181
- goal or a compatible narrowing that changes method or emphasis, not requested
182
- outcomes. A completed-but-unclosed session must close as completed before a
183
- new request proceeds or a new plan is saved. Otherwise a materially new or
184
- expanded request makes no mutation; offer to continue, defer, or abandon the
185
- active session. Keep this a conversational judgment with no classifier or new
186
- state.
122
+ - On top-level error, report exact summary/recovery and, if
123
+ \`workflowData.delivery\` exists, the handoff below. State this initial read
124
+ made no lifecycle, Git, or release mutation; stop.
125
+ - For \`archiveRetry\`, replay the projected \`flow_session_close\` request
126
+ byte-for-byte. Report delivery; if absent, report exact recovery and no map.
127
+ Refresh only if publication is unconfirmed; continue from the confirmed
128
+ projection or stop without saving. This precedes alignment and grants no work.
129
+ - On close conflict, refresh compact and retry only for the same session/goal
130
+ while status permits that kind; never close a replacement.
131
+ - Before mutation, align the projected goal with \`/flow-plan\`; continue only for
132
+ the same goal or a method/emphasis narrowing that preserves every requested
133
+ outcome, and close completed work. For a non-completed session, an
134
+ explicit deferred/abandoned choice calls
135
+ \`flow_session_close\` with compact id/revision, fresh operation id, that kind,
136
+ and optional summary; report delivery, follow exact \`archiveRetry\`, and stop.
137
+ Other new scope makes no mutation; conversationally offer continue, defer, or
138
+ abandon.
139
+ - Delivery handoff: goal; closure kind/summary; progress; per-feature
140
+ ID/title/attempts/latest state/outcome summary/terminal findings; and
141
+ \`reportedArtifacts.latestAttempts\` plus
142
+ \`reportedArtifacts.supersededAttemptsOnly\`, qualified as Flow-reported
143
+ caller-declared artifacts, not an exact/exhaustive Git delta. Map IDs only
144
+ from \`outcomeSummary\`/\`terminalFindings\`: \`verified\` only when proven,
145
+ otherwise \`incomplete\` or explicit \`deferred\`; \`fixed\` needs passing review
146
+ plus current evidence, \`recurring\` current confirmation, \`residual\` a
147
+ confirmed nonblocker, and \`abandoned\` stays the closure kind. Missing history
148
+ is unavailable; never read detail solely for closure or invent it.
187
149
  - If the user asked only for a plan and an approved same-goal session already
188
- exists, load \`flow_status { request: { view: "detail" } }\` once. Report the
189
- immutable active plan and current progress, then stop. Do not save, approve,
190
- or run anything.
191
- - Do not replace an unclosed different goal. Close or finish it explicitly.
150
+ exists, read detail once, report its immutable plan/progress, and stop without
151
+ saving, approving, or running.
192
152
  - If \`flow_plan_save\` or \`flow_plan_approve\` is unavailable, stop and report
193
- that the Flow plugin is not fully loaded.
194
- - Inspect relevant code, tests, docs, package scripts, and local conventions.
195
- Resolve repository facts by inspection; ask the user only when a missing
196
- product choice would materially change the outcome.
153
+ an incomplete plugin load.
154
+ - Inspect relevant code, tests, docs, scripts, and conventions. Ask only for a
155
+ missing product choice that materially changes the outcome.
197
156
 
198
157
  ## Plan contract
199
158
 
@@ -206,27 +165,21 @@ Save one plan with:
206
165
  - \`features\`: ordered outcome slices, each with a stable \`id\`, \`title\`,
207
166
  \`summary\`, bounded \`targets\`, concrete \`validation\`, and \`dependsOn\` ids.
208
167
 
209
- Each feature should have one observable outcome that a reviewer can judge
210
- pass/fail from bounded evidence, plus one focused validation story. Split only
211
- when two outcomes can fail independently or a true dependency requires ordered
212
- acceptance. Keep behavior together when it shares one invariant or neither part
213
- can be accepted alone. Overlapping files by themselves force neither a split nor
214
- a merge. A reviewer should not re-audit the whole product to decide whether one
215
- feature passed. Avoid step-shaped features such as “update files” and vague
216
- checks such as “run tests.”
217
-
218
- When the request or its source material names stable finding, issue, or
219
- requirement IDs, preserve those exact IDs in the relevant saved feature
220
- \`summary\` or \`validation\` prose. Do not replace an ID such as \`F3\`, \`B12\`, or
221
- \`ISSUE-42\` with an unnamed generalization. Every named ID must remain traceable
222
- from the immutable plan to one feature outcome and its evidence.
223
-
224
- When a \`validation\` entry names an executable command, record the exact
225
- plan-listed command byte-for-byte. Behavior-oriented prose that has never run
226
- as an exact command remains reviewer judgment rather than a fabricated command
227
- result. Name any required operating system, architecture, service, credential,
228
- external setting, hardware, or other evidence environment explicitly enough
229
- that \`flow-run\` can preflight it before implementation.
168
+ Each feature needs one observable outcome judgeable from bounded evidence and
169
+ focused validation. Split only independent failures or true dependencies; file
170
+ overlap decides neither. Separate a race or state-machine invariant from
171
+ independently acceptable UI, persistence, or accessibility outcomes; merge only
172
+ under one indivisible invariant. Avoid step-shaped features and vague checks.
173
+
174
+ Preserve stable finding, issue, or requirement IDs exactly in the saved feature
175
+ \`summary\` or \`validation\`; each stays traceable from the immutable plan to one
176
+ outcome and its evidence.
177
+
178
+ When \`validation\` names an executable command, record the exact plan-listed
179
+ command byte-for-byte. Unrun behavior prose remains reviewer judgment, never a
180
+ fabricated command result. Explicitly name any required operating system,
181
+ architecture, service, credential, external setting, hardware, or evidence
182
+ environment for \`flow-run\` to preflight before implementation.
230
183
 
231
184
  Before saving, confirm:
232
185
 
@@ -240,15 +193,14 @@ Before saving, confirm:
240
193
 
241
194
  ## Save and approve
242
195
 
243
- Call \`flow_plan_save\` with one nested request containing a stable operation id,
244
- the current revision (\`0\` for a new session), the goal, and the complete draft.
245
- Summarize the outcome, feature order, validation, and material decisions for
246
- the user. Call \`flow_plan_approve\` with a fresh operation id and current
247
- revision only after explicit approval or when the user already authorized
248
- autonomous implementation. Approval locks the plan. When \`/flow-auto\` needs
249
- conversational approval, ask for it without requiring a second command; a reply
250
- may resume the same process-local auto interaction only after approval advances
251
- the same Flow session.
196
+ Call \`flow_plan_save\` with one nested request: stable operation id, current
197
+ revision (\`0\` for new), goal, and complete draft. Summarize outcome, feature
198
+ order, validation, and material decisions. Call \`flow_plan_approve\` with a fresh
199
+ operation id/current revision only after explicit approval or prior autonomous
200
+ implementation authority. Approval locks the plan. Ask conversational
201
+ \`/flow-auto\` approval without requiring a second command; a reply may resume its
202
+ same process-local interaction only after approval advances the same Flow
203
+ session.
252
204
 
253
205
  Do not begin implementation during a plan-only request. Do not create a plan
254
206
  document in the repository unless the user explicitly requests one.
@@ -277,8 +229,8 @@ bounded packet, assignment-linked validations, approved-plan context, and
277
229
  completed feature IDs instead of reconstructing feature, source,
278
230
  revision, validation, or lifecycle data from conversation memory.
279
231
 
280
- If the reviewer projection is available but evidence required to justify a
281
- verdict is missing, submit a failed result with an ordinary blocking finding
232
+ If the reviewer projection is available but evidence required to approve the
233
+ outcome is missing, submit a failed result with an ordinary blocking finding
282
234
  that precisely identifies the missing evidence.
283
235
  If the assignment itself is unavailable, report that failure without another
284
236
  state change so the manager can inspect compact status. Never invent validation,
@@ -302,6 +254,26 @@ the manager's summary. Check that:
302
254
  - persistence, concurrency, security, migration, compatibility, package, UI,
303
255
  and recovery risks were examined when relevant.
304
256
 
257
+ Finish the supplied feature-specific risk checklist, represented by a bounded
258
+ matrix for concurrency or state-machine work. Continue that matrix after finding
259
+ one blocker so independently detectable interleavings arrive in the same review
260
+ cohort.
261
+
262
+ Scope plan/source IDs by assignment kind. An ordinary feature review records
263
+ dispositions only for IDs mapped to the active feature or explicitly supplied
264
+ in its feature packet; unrelated IDs visible in approved-plan context are
265
+ context, not review claims. A final review traces and records dispositions for
266
+ every approved requirement and feature. Regardless of kind, verify every
267
+ still-live prior disposition against current source and evidence. Terminal \`fixed\`
268
+ requires this review to pass and current evidence to prove the repair. On a
269
+ failed verdict, preserve every prior ID: report a proven repair as
270
+ \`repair proven; terminal fixed pending pass\` with a concise evidence reference
271
+ and carry it into the next attempt. An unproven blocking repair fails under the same ID;
272
+ an unproven advisory repair stays advisory under that ID with its fixed claim
273
+ unverified. Call it \`residual\` only when current evidence confirms the nonblocker
274
+ remains. Escalate only when current evidence makes it outcome-blocking. A
275
+ confirmed blocking recurrence stays blocking under the same ID.
276
+
305
277
  Use the manager-supplied baseline inventory in the assignment for Git-only
306
278
  metadata, and independently inspect the projected changed artifacts with your
307
279
  read-only access. It is evidence, not a verdict. Lack of shell access alone is
@@ -314,27 +286,34 @@ for the delivered state. Flow deliberately projects no raw command output; use
314
286
  the durable command, exit code, completeness, digest, source binding, and your
315
287
  workspace inspection. A weak or unclear coverage claim is an evidence gap.
316
288
 
317
- For a final assignment, also trace every approved requirement and feature to
318
- the delivered result, inspect broad validation, and confirm docs, commands,
319
- package surfaces, and remaining gaps are consistent with completion. The final
320
- assignment is the feature's one review, not a second review layered on top.
321
-
322
- Use \`severity: "blocking"\` only for a concrete issue that invalidates the
323
- approved outcome; otherwise use \`advisory\`. Prefix a blocking finding's summary
324
- with \`[scope-blocker]\` only when resolving it requires material work outside the
325
- approved plan, and identify that boundary in \`evidence\`. No other finding tag
326
- is defined: ordinary in-scope blocking findings and advisory findings need no
327
- tag. Missing evidence is an ordinary, precise blocking finding, not a
328
- \`[scope-blocker]\`.
289
+ For a final assignment, also inspect broad validation and confirm docs,
290
+ commands, package surfaces, and remaining gaps are consistent with completion.
291
+ The final assignment is the feature's one review, not a second review layered
292
+ on top.
293
+
294
+ Start every ordinary finding summary with
295
+ \`finding <feature-id>.R<assignment-createdRevision>-<NN> …\`; for example,
296
+ \`finding frontend-integrity.R12-01 …\`. Reuse a prior ID for recurrence and
297
+ use the current assignment revision plus local sequence for a new issue, so
298
+ dropped history after a qualifying pass cannot cause reuse. Preserve
299
+ source-provided IDs in summary or evidence. Use \`severity: "blocking"\` only
300
+ for a concrete issue that invalidates the approved outcome; otherwise use
301
+ \`advisory\`. Only a blocking issue whose repair requires material work outside
302
+ the approved plan uses the exception grammar
303
+ \`[scope-blocker] finding <feature-id>.R<assignment-createdRevision>-<NN> — …\`;
304
+ identify the boundary in
305
+ \`evidence\`. The marker is forbidden on advisory findings and is the only
306
+ bracketed routing marker. Missing outcome evidence is an ordinary, precise
307
+ blocking finding, not a \`[scope-blocker]\`.
329
308
 
330
309
  Every blocker must map to an approved requirement, changed behavior, or exact
331
310
  missing evidence. Keep its summary precise. In \`evidence\`, cite a changed
332
311
  artifact and location or identify the exact missing evidence or unmet approved
333
312
  requirement.
334
313
 
335
- For missing proof, fail with a precise blocker naming the manager-owned
336
- scenario, command or environment, and expected observable result. Do not pass
337
- conditionally or ask the manager to proxy your verdict.
314
+ For proof missing from the approved outcome, fail with a precise blocker naming
315
+ the manager-owned scenario, command or environment, and expected observable
316
+ result. Do not pass conditionally or ask the manager to proxy your verdict.
338
317
 
339
318
  ## Submit one result
340
319
 
@@ -342,6 +321,18 @@ Call \`flow_feature_complete\` directly with the assignment id, current reviewer
342
321
  projection revision and feature id, a fresh operation id, a concise summary,
343
322
  and exactly one assignment result:
344
323
 
324
+ Keep the summary bounded. For an ordinary review, list as proven \`verified\` or
325
+ \`incomplete\` only plan/source IDs mapped to the active feature or explicitly
326
+ supplied in its feature packet; for a final review, list every approved
327
+ requirement/feature ID. For every still-live supplied prior-finding ID, preserve
328
+ its ID, report current severity and any change from the supplied severity, and
329
+ state confirmed \`recurring\`, confirmed \`residual\` only for a nonblocker, or that
330
+ its fixed claim is unverified. Only a passing result may state proven \`fixed\`. A
331
+ failed result carries every prior ID forward; for a proven repair use
332
+ \`repair proven; terminal fixed pending pass\` plus a concise evidence reference. Copy no
333
+ evidence prose; recurring blockers remain findings. Only IDs fixed by a passing
334
+ review leave the live carry-forward set.
335
+
345
336
  \`\`\`json
346
337
  {
347
338
  "request": {
@@ -383,46 +374,40 @@ description: Implement, validate, independently review, and record one approved
383
374
 
384
375
  # Flow Run
385
376
 
386
- Work on exactly one approved feature. The root manager owns the session,
387
- integration, validation, review dispatch, reset, closure, and every
388
- manager-owned lifecycle mutation. Bounded \`flow-worker\`
389
- instances may contribute disjoint work; the reserved \`flow-reviewer\` owns the
390
- independent review and submits its own result. Never use generic or
391
- general-purpose agents for active Flow work.
377
+ Work on exactly one approved feature.
392
378
 
393
379
  ## Start and scope
394
380
 
395
381
  1. Call \`flow_status { request: { view: "compact" } }\` first. Treat
396
382
  \`nextAction\` as the durable default workflow direction, not as permission.
397
383
  2. If the top-level response status is \`error\`, report its exact summary and
398
- recovery when present. State that this initial read made no lifecycle, Git,
399
- or release mutation, and stop. Do not route an error projection's
400
- \`nextAction\` as feature recovery.
384
+ recovery when present and, if \`workflowData.delivery\` exists, the handoff
385
+ below. State this initial read made no lifecycle, Git, or release mutation;
386
+ stop and never route its \`nextAction\`.
401
387
  3. If compact status contains \`archiveRetry\`, call \`flow_session_close\` once
402
- with that projected request byte-for-byte and report
403
- \`workflowData.delivery\`. If archive publication remains unconfirmed, refresh
404
- compact status. Stop after this cleanup outcome either way; it grants no new
405
- work and therefore precedes goal alignment.
388
+ with the projected request byte-for-byte. Report delivery under the contract
389
+ below. Refresh only if publication is unconfirmed. Stop after this cleanup
390
+ outcome either way; it precedes goal alignment and grants no work.
406
391
  4. When the projection contains an active goal, align it with the current
407
- direct \`/flow-run\` request before any other manager-owned lifecycle mutation.
408
- Continue only for the same goal or a compatible narrowing. Compatible
409
- narrowing may change method or emphasis, but must not add, drop, reorder, or
410
- weaken an approved requirement or feature outcome. A completed-but-unclosed
411
- session must close as completed before a new request proceeds. For other
412
- materially new or expanded work, make no mutation, say the request has not
413
- started, and offer to continue the active goal, defer it, or abandon it.
414
- Keep this comparison conversational; add no classifier or state.
415
- 5. If status is \`idle\` or \`planning\`, report its projected planning action,
392
+ \`/flow-run\` request before another manager lifecycle mutation. Continue only
393
+ for the same goal or a method/emphasis narrowing that preserves all outcomes;
394
+ close completed work. Unless step 5 applies, new/expanded work makes no
395
+ mutation: say it has not started and offer continue, defer, or abandon. Add no
396
+ classifier or state.
397
+ 5. If the aligned request explicitly chooses deferred or abandoned closure for
398
+ a non-completed session, call \`flow_session_close\` with compact session id and
399
+ revision, fresh operation id, that kind, and optional summary. Report delivery
400
+ under the contract below, follow a projected exact \`archiveRetry\`, and stop.
401
+ 6. If status is \`idle\` or \`planning\`, report its projected planning action,
416
402
  explain that \`/flow-run\` requires an approved feature, and stop without
417
403
  mutation.
418
404
 
419
405
  Route every compact projection in this order:
420
406
 
421
- - \`flow_session_close\`: close a completed session with its projected session id
422
- and revision, a fresh operation id, and \`kind: "completed"\`. Report
423
- \`workflowData.delivery\`, follow one projected exact \`archiveRetry\` if needed,
424
- and stop. A materially new request can enter Flow planning afterward; do not
425
- fabricate a run.
407
+ - \`flow_session_close\`: close completed work with its projected session
408
+ id/revision, fresh operation id, and \`kind: "completed"\`. Report delivery
409
+ under the contract below, follow one exact \`archiveRetry\` if needed, and stop.
410
+ New work may enter planning afterward; do not fabricate a run.
426
411
  - \`await-user-direction\` or blocked \`flow_feature_reset\`: call
427
412
  \`flow_status { request: { view: "detail" } }\` exactly once, then distinguish
428
413
  the projected status:
@@ -455,86 +440,92 @@ Route every compact projection in this order:
455
440
  - Any other action: report it and stop unless the runtime explicitly identifies
456
441
  an active execution path.
457
442
 
458
- Use execution status as the active scope and source of revision guards. Read
459
- the feature summary, targets, validation, dependencies, requirements, and
460
- decisions before editing. A projected action may still fail an
461
- environment-sensitive guard; refresh compact status and handle that exact
462
- rejection instead of forcing the stale action.
463
-
464
- Preserve unrelated worktree changes and stay inside the active feature. Leave
465
- changes owned by another planned feature for that feature. If implementation
466
- needs material scope outside the approved plan, stop editing. Finish the
467
- approved plan or have the user explicitly choose deferred or abandoned closure
468
- before starting a new plan; never replan the active approved session in place.
469
- Use \`flow_feature_reset\` when a wrong design or invalid assumption requires a
470
- fresh run within the active feature; do not layer a retry onto a bad execution.
443
+ Use execution status for active scope/revision guards. Before editing, read the
444
+ feature summary, targets, validation, dependencies, requirements, and decisions.
445
+ If a projected action fails an environment-sensitive guard, refresh compact and
446
+ handle that rejection; never force it.
447
+
448
+ When resuming attempt 2 or later and prior findings are not already available,
449
+ read detail once and recover still-live IDs from superseded attempts' latest
450
+ relevant failure before preflight.
451
+
452
+ Summaries keep plan/source IDs \`verified\` or \`incomplete\`. A prior finding
453
+ \`fixed\` needs passing review plus current evidence. A failed review carries
454
+ every prior ID forward: use \`terminal fixed pending pass\` plus concise evidence
455
+ for a proven repair, unverified-fixed for an unproven repair, \`recurring\` when
456
+ confirmed, or \`residual\` for a confirmed nonblocker. Blockers remain terminal.
457
+ Delivery handoff: goal; closure kind/summary; progress; per-feature
458
+ ID/title/attempts/latest state/outcome summary/terminal findings; and
459
+ \`reportedArtifacts.latestAttempts\` plus
460
+ \`reportedArtifacts.supersededAttemptsOnly\`, qualified as Flow-reported
461
+ caller-declared artifacts, not an exact/exhaustive Git delta. Map IDs only from
462
+ delivery \`outcomeSummary\`/\`terminalFindings\`; requirements are \`verified\`,
463
+ \`incomplete\`, or explicitly \`deferred\`, and \`abandoned\` remains the kind.
464
+ If delivery is absent, report exact recovery and no map; never invent or read
465
+ detail solely for closure. On revision conflict, refresh compact; retry only for
466
+ the same session and goal while status still permits the selected closure kind;
467
+ never close a replacement.
468
+
469
+ Preserve unrelated work and stay inside the feature. Out-of-plan work stops;
470
+ finish or obtain explicit deferred/abandoned closure before a new plan. Never
471
+ replan in place. Use \`flow_feature_reset\` for a wrong design or assumption; do
472
+ not layer retries.
471
473
 
472
474
  ## Evidence and risk preflight
473
475
 
474
476
  Before editing or dispatching a worker, perform one preflight from the approved
475
477
  feature and current worktree:
476
478
 
477
- - Preserve every named finding or requirement ID from the feature prose and
478
- map it to an observable acceptance outcome.
479
- - Inventory exact commands and behavior evidence, including any required
480
- operating system, architecture, tool, service, credential, external setting,
481
- or hardware. Confirm an available, authorized path for each.
482
- - Inspect the baseline and unrelated work, including deletions, renames, file
483
- types, and executable modes.
484
- - Write one concise adversarial checklist covering the outcome, failure and
485
- cleanup ordering, adjacent states, repetition/retry/interruption/concurrency,
486
- overlapping invariants, and other relevant platform or persistence risks.
487
-
488
- Use the checklist and named IDs in every worker assignment and the review
489
- packet. If required evidence needs user or external authority, stop before
490
- implementation and ask. While required behavior or environment evidence is
491
- knowingly skipped or unavailable, manager policy forbids calling
492
- \`flow_review_start\`; a substitute pass does not cure the gap. Flow persists no
493
- skipped-evidence ledger, and the reviewer treats missing proof as blocking if
494
- the gap reaches its packet.
479
+ - Preserve every named finding/requirement and still-live prior review ID; map
480
+ each to an observable acceptance outcome.
481
+ - Inventory exact commands, behavior evidence, required operating system,
482
+ architecture, service, credential, external setting, or hardware, and an
483
+ authorized path for each.
484
+ - Reuse one conversational run baseline of unrelated work, deletions, renames,
485
+ file types, and executable modes. Refresh changed facts; give each review only
486
+ facts the feature changes or depends on, and give final review the full
487
+ inventory.
488
+ - Write one concise adversarial checklist covering failure and cleanup ordering,
489
+ adjacent states, repetition, retry, interruption, concurrency, overlapping
490
+ invariants, and relevant platform or persistence risks. For concurrency or
491
+ state-machine work, express it as a compact matrix with \`state/interleaving\`,
492
+ \`event\`, \`expected outcome\`, \`cleanup/invariant\`, and \`evidence\` columns.
493
+
494
+ Carry the checklist/IDs through workers and review; on retry preserve still-live
495
+ prior IDs/dispositions. Required evidence needing user or external authority
496
+ stops before implementation. If skipped or unavailable, it forbids
497
+ \`flow_review_start\`; a substitute pass cannot cure it. Flow persists no
498
+ skipped-evidence ledger; the reviewer blocks the gap.
495
499
 
496
500
  ## Implement
497
501
 
498
- Prefer the smallest change that satisfies the approved outcome. Follow the
499
- repository's existing boundaries and conventions. Do not create lifecycle,
500
- validation, audit, or handoff sidecars. Durable user-requested reports should
501
- normally be one stable Markdown artifact; JSON requires an explicit request.
502
+ Make the smallest change satisfying the approved outcome and repository
503
+ boundaries. Create no lifecycle, validation, audit, or handoff sidecars. A
504
+ durable user-requested report is normally one stable Markdown artifact; JSON
505
+ requires an explicit request.
502
506
 
503
- Do not stage, commit, push, publish, or mutate releases unless the user asks for
504
- that separate action.
507
+ Do not stage, commit, push, publish, or mutate releases unless asked separately.
505
508
 
506
509
  ## Bounded worker waves
507
510
 
508
- Work serially by default. Existing implementation authority covers a qualifying
509
- worker wave; do not ask for separate approval. After manager orientation, fan
510
- out only when two or three genuinely independent, non-overlapping slices can be
511
- named and parallel execution has clear benefit. Run one cohort of two or three
512
- \`flow-worker\` instances at a time. Issue every cohort Task call in the same
513
- assistant tool-use turn before consuming any result. If the host or model
514
- serializes those calls, treat and report that execution as serial instead of
515
- claiming parallelism. Each prompt must name a stable slice id, the exact outcome
516
- and read or write scope, expected coverage, recommended manager checks,
517
- dependencies, a stop condition, and the applicable adversarial acceptance and
518
- risk checklist from preflight. A worker must receive the checklist before it
519
- codes. Edit scopes must be exact and non-overlapping. Shared contracts,
520
- lockfiles, and generated outputs remain manager-owned unless one worker
521
- receives the whole relevant scope. Never substitute a generic agent for
522
- \`flow-worker\`, including for read-only evidence gathering.
523
-
524
- Workers cannot call Flow tools or spawn children. Each returns one concise
525
- handoff containing status, scope and coverage, evidence or changed paths,
526
- recommended manager checks, gaps and risks, and integration notes. Workers do
527
- not run Bash; all executable checks remain manager-owned. Missing, partial, or
528
- blocked output remains an explicit coverage gap.
529
-
530
- After all workers stop, compare actual changed paths with every assigned scope,
531
- then inspect the combined diff and evidence and reconcile unexpected paths or
532
- conflicts before validation. At most one targeted follow-up wave may address a
533
- failed slice, newly unlocked dependency, or material claim verification. Do
534
- not start an automatic third wave. Coordination stays in the conversation:
535
- create no manifest, sidecar, Session field, durable handoff, or recovery ledger.
536
- After an interruption, inspect Flow status and the worktree and treat partial
537
- worker edits as untrusted.
511
+ Work serially by default; existing authority covers a qualifying worker wave.
512
+ After manager orientation, fan out only two or three genuinely independent,
513
+ non-overlapping slices with clear benefit. Issue one cohort's Task calls in the
514
+ same assistant tool-use turn; report host serialization as serial. Each
515
+ assignment names a stable id, exact outcome/read-write scope, coverage, manager
516
+ checks, dependencies, stop condition, and preflight risk checklist. The worker
517
+ must receive the checklist before it codes. Shared contracts, lockfiles, and generated
518
+ output stay manager-owned unless wholly assigned to one worker.
519
+
520
+ Workers call no Flow tools, spawn no children, and run no Bash. Each returns
521
+ status, scope/coverage, evidence/changed paths, manager checks, gaps/risks, and
522
+ integration notes; missing or blocked output is a coverage gap.
523
+
524
+ After workers stop, reconcile paths/scopes and inspect combined diff/evidence
525
+ before validation. At most one targeted follow-up wave may repair a slice,
526
+ unlock a dependency, or verify a material claim; never a third. Create no
527
+ coordination ledger/sidecar. After interruption, inspect status/worktree and
528
+ treat partial worker edits as untrusted.
538
529
 
539
530
  ## Validate
540
531
 
@@ -552,15 +543,13 @@ checks from the changed behavior and risk:
552
543
  for the repository's canonical applicable gate or a justified equivalent
553
544
  that covers the delivered repository state.
554
545
 
555
- Immediately before each Bash command used as evidence, call
556
- \`flow_validation_start\` with the current revision, feature id, the exact
557
- command, and \`scope\` (\`focused\` or \`broad\`). Run that byte-for-byte command next
558
- and inspect its complete outcome. Flow records the host-observed result directly
559
- in the session; do not copy host-observed fields into a later request. The exact
560
- command is durable, so never inline tokens, passwords, credentials, or other
561
- secrets. Raw output is deliberately neither persisted nor projected: the
562
- durable evidence is the command, exit code, output completeness, and output
563
- digest, while the manager must inspect the live output.
546
+ Immediately before each evidence Bash command, call \`flow_validation_start\`
547
+ with current revision, feature id, exact command, and \`scope\` (\`focused\` or
548
+ \`broad\`). Run it byte-for-byte next and inspect the complete outcome. Flow
549
+ records the host observation; copy no host-observed fields into a later request.
550
+ The command is durable, so include no secrets. Raw output is neither persisted
551
+ nor projected: durable evidence is command, exit code, completeness, and digest;
552
+ the manager inspects live output.
564
553
 
565
554
  Exact plan-listed gate commands are recorded byte-for-byte.
566
555
  A failed, incomplete, or source-drifted exact plan-listed observation creates a
@@ -609,26 +598,29 @@ changed no repository artifact. Flow selects current applicable validation
609
598
  automatically and derives \`feature\` versus \`final\` review from plan progress;
610
599
  callers do not supply the review kind.
611
600
 
612
- The packet must preserve the feature's named finding and requirement IDs and
613
- summarize the preflight checklist, adjacent state transitions, repeated and
614
- failure-path behavior, overlapping feature invariants, and the inspected base
615
- diff including deletions, renames, file types, and executable-mode changes.
616
- Call out any item that needs independent reviewer scrutiny; do not hide a known
617
- evidence gap in prose.
618
-
619
- Dispatch the returned assignment only to the reserved \`flow-reviewer\`. Do not
620
- perform the independent review in manager context and never copy or submit its
621
- verdict. The reviewer reads its assignment, inspects the workspace, and calls
622
- \`flow_feature_complete\` directly; the runtime verifies the calling agent. The
623
- reviewer remains workspace-read-only and may make only this exact result
624
- submission as its sole lifecycle mutation.
625
-
626
- After dispatch, read compact status rather than trusting reviewer prose. If the
627
- top-level response is an error, report its exact summary and recovery when
628
- present, say the latest lifecycle state could not be confirmed, and stop
629
- without further mutation. Do not claim this invocation made no lifecycle
630
- mutation: it may already have started review or recorded a reviewer result.
631
- Never invent or submit a verdict. If status remains running, apply the
601
+ Keep the packet bounded. Map IDs to current-source commands or scenarios,
602
+ environment, and results. Put the feature-specific risk checklist under
603
+ \`Risks/Matrix:\`, representing it as a transition matrix for concurrency or
604
+ state-machine work. Include \`Baseline:\` facts only when this feature changes or
605
+ depends on them, except that final review receives the full inventory. Under
606
+ \`Prior findings:\`, preserve each still-live prior ID's original severity and
607
+ current \`terminal fixed pending pass\`, \`recurring\`, \`residual\`, or
608
+ unverified-fixed state. Ordinary-review plan/source IDs are limited to
609
+ active-feature mappings or IDs explicitly supplied for its packet; final review
610
+ includes every approved requirement/feature ID. Always carry still-live prior
611
+ findings. Omit empty optional sections; state \`none\` only for a relevant
612
+ inspected absence. Never hide a gap.
613
+
614
+ Dispatch only to reserved \`flow-reviewer\`; never review, copy, or submit its
615
+ verdict in manager context. It reads the assignment/workspace and calls
616
+ \`flow_feature_complete\` directly; runtime verifies the caller. It stays
617
+ workspace-read-only, with that result submission as its sole lifecycle mutation.
618
+
619
+ After dispatch, read compact status. On top-level error, report exact
620
+ summary/recovery, say the latest lifecycle state could not be confirmed, and
621
+ stop without further mutation. Do not claim this invocation made no lifecycle
622
+ mutation: review may have started or recorded a result. Never invent or submit a
623
+ verdict. If status remains running, apply the
632
624
  \`dispatch-flow-reviewer\` or running \`flow_feature_reset\` route above. If status
633
625
  is blocked, load detail through the single blocked route above. A recorded pass
634
626
  completes the feature.
@@ -638,10 +630,11 @@ completes the feature.
638
630
  Use compact \`blockedFeature.failedReviewCount\` with the one detail projection.
639
631
 
640
632
  - A \`[scope-blocker]\` checkpoints immediately. Do not reset automatically.
641
- - On the first ordinary failed review, existing implementation authority
642
- permits one automatic \`flow_feature_reset\` with the blocked \`featureId\` as
643
- \`nextFeatureId\`. That call atomically starts the fresh full retry. Fix only its
644
- blocking findings, then run full validation and full independent review.
633
+ - Only when \`failedReviewCount === 1\` and there is no \`[scope-blocker]\`,
634
+ existing implementation authority permits one automatic \`flow_feature_reset\`
635
+ with the blocked \`featureId\` as \`nextFeatureId\`. That call atomically starts
636
+ the fresh full retry. Fix only its blocking findings, then run full validation
637
+ and full independent review.
645
638
  - A feature whose latest relevant reviewed outcome remains failed is never
646
639
  selected implicitly. \`/flow-auto\` may still continue an untouched,
647
640
  dependency-independent feature. When every runnable candidate requires a
@@ -649,7 +642,7 @@ Use compact \`blockedFeature.failedReviewCount\` with the one detail projection.
649
642
  has already been superseded: after explicit direction, read detail once and
650
643
  call \`flow_run_start\` with the exact retry feature ID. Do not reset from that
651
644
  ready checkpoint.
652
- - After the second failed review, retry only when the current aligned request
645
+ - When \`failedReviewCount >= 2\`, retry only when the current aligned request
653
646
  explicitly authorizes one additional attempt. Pass the blocked feature as
654
647
  \`nextFeatureId\` on \`flow_feature_reset\`; if that attempt fails, checkpoint
655
648
  again.
@@ -659,13 +652,12 @@ Use compact \`blockedFeature.failedReviewCount\` with the one detail projection.
659
652
  attempts and starts that exact run in one transaction. Do not reset first,
660
653
  call \`flow_run_start\` separately, or rely on default selection.
661
654
 
662
- When stopping blocked, label the result overall incomplete. Report what the
663
- latest repair fixed; recurring and new blocking findings; the goal and progress;
664
- the blocked feature, attempt, and failure count; completed and untouched
665
- features; latest validations and \`artifactsChanged\` as Flow-reported artifact
666
- evidence; Git and release mutation status; whether the current request started
667
- and matched the active goal; the exact \`nextAction\`; and whether another attempt
668
- requires explicit authorization.
655
+ When stopping blocked, label overall incomplete. Report the latest repair proved
656
+ pending a passing review; recurring and new blockers; goal/progress; blocked
657
+ feature, attempt, and failure count; completed/untouched features; latest
658
+ validations and \`artifactsChanged\` as Flow-reported artifact evidence; Git/release
659
+ mutation status; whether this request started and matched the goal; exact
660
+ \`nextAction\`; and whether another attempt requires explicit authorization.
669
661
 
670
662
  Use that already-loaded compact status after every recorded outcome. Direct
671
663
  \`/flow-run\` reports this one feature's cumulative outcome and \`nextAction\`, then
@@ -682,10 +674,38 @@ var FLOW_GUIDANCE_TOPICS = [
682
674
  var FLOW_GUIDANCE_IDS = FLOW_GUIDANCE_TOPICS;
683
675
 
684
676
  // src/guidance/catalog.ts
677
+ var FLOW_MANAGER_KERNEL = [
678
+ "## Flow manager kernel",
679
+ "",
680
+ [
681
+ "- The root manager owns manager lifecycle mutations, integration, validation, and review dispatch;",
682
+ "the independent reviewer submits only its own result."
683
+ ].join(" "),
684
+ [
685
+ "- Delegate active Flow work only to `flow-worker` and independent review only to `flow-reviewer`;",
686
+ "never use generic or general-purpose agents."
687
+ ].join(" "),
688
+ [
689
+ "- Make one automatic fresh full retry only when `failedReviewCount === 1`",
690
+ "and no `[scope-blocker]` is present; otherwise checkpoint."
691
+ ].join(" "),
692
+ [
693
+ "- Before review, require current-source evidence appropriate to the changed outcome,",
694
+ "including behavior evidence when behavior changes, plus relevant base-diff, deletion,",
695
+ "rename, file-type, and executable-mode facts."
696
+ ].join(" ")
697
+ ].join(`
698
+ `);
685
699
  var GUIDANCE_CONTENT = {
686
- flow: SKILL_default,
700
+ flow: `${SKILL_default.trimEnd()}
701
+
702
+ ${FLOW_MANAGER_KERNEL}
703
+ `,
687
704
  "flow-plan": SKILL_default2,
688
- "flow-run": SKILL_default4,
705
+ "flow-run": `${SKILL_default4.trimEnd()}
706
+
707
+ ${FLOW_MANAGER_KERNEL}
708
+ `,
689
709
  "flow-review": SKILL_default3
690
710
  };
691
711
  var FLOW_GUIDANCE_DOCUMENTS = FLOW_GUIDANCE_TOPICS.map((name) => ({
@@ -706,47 +726,99 @@ if (FLOW_GUIDANCE_BY_ID.size !== FLOW_GUIDANCE_IDS.length) {
706
726
  }
707
727
 
708
728
  // src/prompt-surfaces.ts
709
- var FLOW_WORKER_PROMPT = `# Flow bounded worker
710
-
711
- You are one hidden Flow worker supporting the root manager inside one active feature. Own only the single slice explicitly assigned by that manager. Preserve all unrelated work and do not broaden the assignment.
712
-
713
- You may run concurrently with sibling workers. Do not enter their scopes, assume their results, or revert changes you did not make.
714
-
715
- ## Scope and authority
716
-
717
- - Use the manager assignment as your only source of Flow lifecycle context. Do not call any \`flow_*\` tool, including \`flow_status\`.
718
- - Do not delegate, spawn subtasks, or load skills.
719
- - Do not stage, commit, push, publish, or create a release.
720
- - Do not run Bash commands. The manager owns every executable check.
721
- - Never edit .flow or .git metadata paths; the host denies those paths.
722
- - A read-only evidence slice must not edit files.
723
- - The assignment must include an adversarial acceptance and risk checklist prepared before coding. If it is missing, stop without editing and report the gap.
724
- - An implementation slice may edit only the exact, non-overlapping write paths explicitly assigned by the manager. If required work would escape those paths, stop and return a partial or blocked handoff instead of expanding scope.
725
- - Use only non-shell inspection relevant to the assigned slice. The manager owns integration, focused checks, and authoritative combined validation after all workers have stopped.
726
-
727
- Before editing, apply the supplied checklist to primary behavior, failure and cleanup ordering, adjacent state transitions, repeated or interrupted operation, overlapping invariants, and relevant persistence, concurrency, security, compatibility, or file-metadata risks. Preserve every named finding or requirement ID in your handoff.
728
-
729
- ## Handoff
730
-
731
- Return exactly one concise handoff using this structure:
732
-
733
- ## Status
734
- success | partial | blocked
735
-
736
- ## Scope & coverage
737
- - Assigned slice and what was covered
738
-
739
- ## Findings / changed paths
740
- - Evidence found or exact paths changed
741
-
742
- ## Recommended manager checks
743
- - Exact checks the manager should run, or none
744
-
745
- ## Gaps & risks
746
- - Missing coverage, blockers, conflicts, or none
729
+ var FLOW_WORKER_PROMPT = [
730
+ "# Flow bounded worker",
731
+ [
732
+ "You are one hidden Flow worker supporting the root manager inside one active feature.",
733
+ "Own only the assigned slice, preserve unrelated work, and do not broaden it or enter a sibling's scope."
734
+ ].join(" "),
735
+ "## Scope and authority",
736
+ [
737
+ "- Use the manager assignment as your only source of Flow lifecycle context.",
738
+ "Do not call any `flow_*` tool, including `flow_status`."
739
+ ].join(" "),
740
+ "- Do not delegate, spawn subtasks, or load skills.",
741
+ "- Do not stage, commit, push, publish, or create a release.",
742
+ "- Do not run Bash commands. The manager owns every executable check.",
743
+ "- Never edit .flow or .git metadata paths; the host denies those paths.",
744
+ "- A read-only evidence slice must not edit files.",
745
+ [
746
+ "- The assignment must include an adversarial acceptance and risk checklist,",
747
+ "represented as a transition matrix for concurrency or state-machine work, prepared before coding.",
748
+ "If it is missing, stop without editing and report the gap."
749
+ ].join(" "),
750
+ [
751
+ "- An implementation slice may edit only the exact, non-overlapping write paths explicitly assigned by the manager.",
752
+ "If required work would escape those paths, stop and return a partial or blocked handoff",
753
+ "instead of expanding scope."
754
+ ].join(" "),
755
+ [
756
+ "- The manager owns integration, focused checks, and authoritative combined validation",
757
+ "after all workers have stopped."
758
+ ].join(" "),
759
+ [
760
+ "Before editing, apply the supplied risk coverage through its matrix rows when present:",
761
+ "primary behavior, failure and cleanup ordering, adjacent state transitions,",
762
+ "repeated or interrupted operation, overlapping invariants, and relevant persistence,",
763
+ "concurrency, security, compatibility, or file-metadata risks.",
764
+ "Preserve every named finding, requirement, or prior review ID in your handoff."
765
+ ].join(" "),
766
+ "## Handoff",
767
+ [
768
+ "Return exactly one concise handoff with `Status` (success, partial, or blocked),",
769
+ "`Scope & coverage`, `Findings / changed paths`, `Recommended manager checks`,",
770
+ "`Gaps & risks`, and `Integration notes`."
771
+ ].join(" ")
772
+ ].join(`
773
+ `);
774
+ var FLOW_STATUS_PROMPT = [
775
+ 'Call `flow_status { request: { view: "compact" } }` first.',
776
+ "Do not mutate.",
777
+ "If the top-level response status is `error`, report its exact summary and",
778
+ "`workflowData.failure.recovery` when present; otherwise say no recovery guidance was supplied.",
779
+ "When `workflowData.delivery` is present, also report its goal; closure kind and summary;",
780
+ "progress; for every feature, `id`, `title`, `attempts`, `latestState`, `outcomeSummary`,",
781
+ "and `terminalFindings`; and `reportedArtifacts.latestAttempts` plus",
782
+ "`reportedArtifacts.supersededAttemptsOnly`, qualified as Flow-reported caller-declared artifacts,",
783
+ "not an exact or exhaustive Git delta.",
784
+ "For the terminal ID map use only `outcomeSummary` and `terminalFindings`: IDs are `verified`",
785
+ "only when proven, otherwise `incomplete` or explicitly `deferred`; `fixed` needs later passing",
786
+ "review plus current evidence, `recurring` current confirmation, `residual` a confirmed nonblocker,",
787
+ "and `abandoned` remains the closure kind.",
788
+ "Missing IDs are unavailable.",
789
+ "State that `/flow-status` made no Git or release mutation, report any lifecycle state effect",
790
+ "disclosed by the response, and stop.",
791
+ "Do not interpret recovery guidance as a blocked review.",
792
+ "If `projection.status` is `blocked` or `projection.nextAction` is `await-user-direction`,",
793
+ 'call `flow_status { request: { view: "detail" } }` exactly once and label the result overall incomplete.',
794
+ "From that detail projection, report the goal and progress; any blocked feature, attempt,",
795
+ "`failedReviewCount`, and findings; every retry-required feature whose latest relevant reviewed",
796
+ "outcome remains failed; completed and untouched features; validations and `artifactsChanged`",
797
+ "as Flow-reported artifact evidence; and the exact status and `nextAction`.",
798
+ "For a blocked first failed review, explain that `flow_feature_reset` is only the default and",
799
+ "`/flow-run` must inspect any `[scope-blocker]` before reset.",
800
+ "For blocked `await-user-direction`, explain that an authorized retry or independent choice",
801
+ "uses atomic `flow_feature_reset` with `nextFeatureId`.",
802
+ "For ready `await-user-direction`, explain that no blocked run remains, so an authorized retry",
803
+ "uses `flow_run_start` with an explicit `featureId`, never reset or default selection.",
804
+ "When `workflowData.autoTiming` is present, report `activeMs` as non-authoritative process-local",
805
+ "wall time classified active, not CPU or pure work, and `waitingForUserMs` as only projected",
806
+ "`flow_plan_approve` plus `await-user-direction` time for the latest `/flow-auto`.",
807
+ "State that paused, inactive, errored, and unprojected waits are excluded.",
808
+ "Otherwise report the compact projection and its exact `nextAction`, state that `/flow-status`",
809
+ "made no lifecycle, Git, or release mutation, and stop."
810
+ ].join(" ");
811
+ var FLOW_REVIEW_PROMPT = [
812
+ "# Flow review command",
813
+ [
814
+ "Run this assignment only as the reserved `flow-reviewer`.",
815
+ "The reviewer is independent and workspace-read-only; it may read reviewer status",
816
+ "and submit only its own result through `flow_feature_complete`."
817
+ ].join(" "),
818
+ "Assignment: $ARGUMENTS"
819
+ ].join(`
747
820
 
748
- ## Integration notes
749
- - What the manager must verify or integrate, or none`;
821
+ `);
750
822
  function skillBody(id) {
751
823
  return getFlowGuidance(id).content.replace(/^---\r?\n[\s\S]*?\r?\n---\r?\n/, "").trim();
752
824
  }
@@ -779,26 +851,9 @@ function compileFlowPromptSurface(surface) {
779
851
  case "flow-run":
780
852
  return managerCommand(surface);
781
853
  case "flow-status":
782
- return [
783
- 'Call `flow_status { request: { view: "compact" } }` first.',
784
- "Do not mutate.",
785
- "If the top-level response status is `error`, report its exact summary and `workflowData.failure.recovery` when present; otherwise say no recovery guidance was supplied. State that `/flow-status` made no Git or release mutation, report any lifecycle state effect disclosed by the response, and stop. Do not interpret recovery guidance as a blocked review.",
786
- 'If `projection.status` is `blocked` or `projection.nextAction` is `await-user-direction`, call `flow_status { request: { view: "detail" } }` exactly once and label the result overall incomplete.',
787
- "From that detail projection, report the goal and progress; any blocked feature, attempt, `failedReviewCount`, and findings; every retry-required feature whose latest relevant reviewed outcome remains failed; completed and untouched features; validations and `artifactsChanged` as Flow-reported artifact evidence; and the exact status and `nextAction`.",
788
- "For a blocked first failed review, explain that `flow_feature_reset` is only the default and `/flow-run` must inspect any `[scope-blocker]` before reset.",
789
- "For blocked `await-user-direction`, explain that an authorized retry or independent choice uses atomic `flow_feature_reset` with `nextFeatureId`. For ready `await-user-direction`, explain that no blocked run remains, so an authorized retry uses `flow_run_start` with an explicit `featureId`, never reset or default selection.",
790
- "When `workflowData.autoTiming` is present, report `activeMs` as non-authoritative process-local wall time classified active, not CPU or pure work, and `waitingForUserMs` as only projected `flow_plan_approve` plus `await-user-direction` time for the latest `/flow-auto`. State that paused, inactive, errored, and unprojected waits are excluded.",
791
- "Otherwise report the compact projection and its exact `nextAction`, state that `/flow-status` made no lifecycle, Git, or release mutation, and stop."
792
- ].join(" ");
854
+ return FLOW_STATUS_PROMPT;
793
855
  case "flow-review":
794
- return [
795
- "# Flow review command",
796
- "",
797
- "Run this assignment only as the reserved `flow-reviewer`. The reviewer is independent and workspace-read-only; it may read reviewer status and submit only its own result through `flow_feature_complete`.",
798
- "",
799
- "Assignment: $ARGUMENTS"
800
- ].join(`
801
- `);
856
+ return FLOW_REVIEW_PROMPT;
802
857
  case "flow-reviewer":
803
858
  return skillBody("flow-review");
804
859
  case "flow-worker":
@@ -3479,20 +3534,37 @@ function resolveFlowPluginVersion() {
3479
3534
 
3480
3535
  // src/platform/opencode/auto-drive.ts
3481
3536
  var FLOW_AUTO_METADATA_KEY = "opencode-plugin-flow/auto";
3482
- function continuationToken(parts) {
3537
+ var STOP = /^(?:(?:stop|cancel) \/flow-auto|\/flow-auto (?:stop|cancel))$/i;
3538
+ var CONTINUATION_ROUTE = [
3539
+ "Load flow-run guidance before any feature or closure route;",
3540
+ "for a fresh close use compact session id/revision plus a fresh operation id,",
3541
+ "and replay archiveRetry exactly from its projected request."
3542
+ ].join(" ");
3543
+ function inspectMessage(parts) {
3544
+ let token = null;
3545
+ let text = "";
3546
+ let user = false;
3483
3547
  for (const part of parts) {
3484
- const token = part.metadata?.[FLOW_AUTO_METADATA_KEY];
3485
- if (part.synthetic === true && typeof token === "string")
3486
- return token;
3548
+ if (part.synthetic === true) {
3549
+ const value = part.metadata?.[FLOW_AUTO_METADATA_KEY];
3550
+ if (typeof value === "string")
3551
+ token = value;
3552
+ } else {
3553
+ user = true;
3554
+ text += ` ${part.text ?? ""}`;
3555
+ }
3487
3556
  }
3488
- return null;
3557
+ return { token, user, text: text.trim().replace(/\s+/g, " ") };
3489
3558
  }
3490
3559
  function isMechanical(projection) {
3491
- return projection.status === "ready" && projection.nextAction === "flow_run_start" || (projection.status === "completed" || projection.status === "closed") && projection.nextAction === "flow_session_close";
3560
+ return projection.nextAction === "flow_run_start" ? projection.status === "ready" : projection.nextAction === "flow_session_close" && (projection.status === "completed" || projection.status === "closed");
3492
3561
  }
3493
3562
  function isCheckpoint(projection) {
3494
3563
  return ["flow_plan_approve", "await-user-direction"].includes(projection.nextAction ?? "");
3495
3564
  }
3565
+ function isPendingReviewer(projection) {
3566
+ return projection.status === "running" && projection.nextAction === "dispatch-flow-reviewer";
3567
+ }
3496
3568
 
3497
3569
  class AutoDriveCoordinator {
3498
3570
  #lease = null;
@@ -3525,16 +3597,26 @@ class AutoDriveCoordinator {
3525
3597
  #stop(lease, warning) {
3526
3598
  if (this.#lease !== lease)
3527
3599
  return;
3528
- this.#lease = null;
3529
- this.#setTiming("inactive");
3600
+ this.deactivate(lease.hostSessionId);
3530
3601
  if (warning)
3531
3602
  this.#warn(warning);
3532
3603
  }
3604
+ #rejectOrigin(lease, kind) {
3605
+ this.#stop(lease, `Flow: ${kind} origin was unavailable.`);
3606
+ }
3607
+ #waitAt(lease, revision) {
3608
+ const current = lease.checkpoint;
3609
+ lease.checkpoint = current?.revision === revision ? current : { revision, answered: false };
3610
+ lease.checkpoint.answered = false;
3611
+ lease.messageId = null;
3612
+ lease.lastPromptedRevision = null;
3613
+ this.#setTiming("waiting-for-user");
3614
+ }
3533
3615
  async#read(lease) {
3534
3616
  try {
3535
3617
  return await this.#options.readProjection();
3536
3618
  } catch (error) {
3537
- this.#stop(lease, `Flow auto-drive stopped because compact status failed: ${error instanceof Error ? error.message : String(error)}`);
3619
+ this.#stop(lease, `Flow auto status failed: ${String(error)}`);
3538
3620
  return null;
3539
3621
  }
3540
3622
  }
@@ -3550,12 +3632,16 @@ class AutoDriveCoordinator {
3550
3632
  hostSessionId,
3551
3633
  token,
3552
3634
  baseline: null,
3553
- flowSessionId: null,
3554
3635
  delivery: null,
3555
3636
  lastPromptedRevision: null,
3556
3637
  checkpoint: null,
3557
- pendingReply: null,
3558
- inFlight: null
3638
+ pendingReply: false,
3639
+ inFlight: null,
3640
+ idlePending: false,
3641
+ messageId: null,
3642
+ assistantParents: new Map,
3643
+ lastAssistantParent: null,
3644
+ compaction: null
3559
3645
  };
3560
3646
  const lease = this.#lease;
3561
3647
  const baseline = await this.#read(lease);
@@ -3563,15 +3649,15 @@ class AutoDriveCoordinator {
3563
3649
  throw new Error("Flow auto-drive compact status failed.");
3564
3650
  lease.baseline = baseline;
3565
3651
  if (this.#lease !== lease)
3566
- throw new Error("Flow auto-drive activation superseded.");
3567
- lease.flowSessionId = lease.baseline.sessionId ?? null;
3652
+ throw new Error("Flow auto-drive superseded.");
3653
+ if (!isPendingReviewer(baseline))
3654
+ lease.checkpoint = { revision: baseline.revision, answered: false };
3568
3655
  return { [FLOW_AUTO_METADATA_KEY]: token };
3569
3656
  }
3570
3657
  deactivate(hostSessionId) {
3571
3658
  if (this.#lease?.hostSessionId !== hostSessionId)
3572
3659
  return false;
3573
- this.#lease = null;
3574
- this.#setTiming("inactive");
3660
+ this.clear();
3575
3661
  return true;
3576
3662
  }
3577
3663
  clear() {
@@ -3579,38 +3665,115 @@ class AutoDriveCoordinator {
3579
3665
  this.#setTiming("inactive");
3580
3666
  this.#lease = null;
3581
3667
  }
3582
- async observeMessage(hostSessionId, delivery, parts) {
3668
+ async observeMessage(hostSessionId, delivery, parts, messageId) {
3583
3669
  const lease = this.#lease;
3584
- const token = continuationToken(parts);
3585
- if (token !== null) {
3586
- if (!lease || lease.hostSessionId !== hostSessionId || lease.token !== token) {
3670
+ const message = inspectMessage(parts);
3671
+ if (lease?.hostSessionId === hostSessionId && STOP.test(message.text)) {
3672
+ this.deactivate(hostSessionId);
3673
+ return "accepted";
3674
+ }
3675
+ if (message.token !== null) {
3676
+ if (!lease || lease.hostSessionId !== hostSessionId || lease.token !== message.token)
3587
3677
  return "stale-continuation";
3588
- }
3589
3678
  lease.delivery = delivery;
3679
+ lease.messageId = messageId;
3680
+ if (!lease.checkpoint && lease.lastPromptedRevision !== null)
3681
+ lease.checkpoint = {
3682
+ revision: lease.lastPromptedRevision,
3683
+ answered: false
3684
+ };
3590
3685
  this.#setTiming("active");
3591
- } else if (lease?.hostSessionId === hostSessionId && !parts.every((part) => part.synthetic === true)) {
3592
- if (lease.inFlight === "status") {
3593
- lease.pendingReply = delivery;
3594
- } else if (!lease.checkpoint) {
3595
- this.deactivate(hostSessionId);
3596
- } else {
3597
- const projection = await this.#read(lease);
3598
- if (!projection)
3599
- return "accepted";
3600
- if (this.#lease !== lease || projection.sessionId !== lease.flowSessionId) {
3601
- this.#stop(lease);
3602
- return "accepted";
3603
- }
3604
- lease.checkpoint.revision = Math.max(lease.checkpoint.revision, projection.revision);
3605
- lease.checkpoint.answered = true;
3606
- lease.delivery = delivery;
3607
- this.#setTiming("active");
3608
- }
3686
+ return "accepted";
3687
+ }
3688
+ if (lease?.hostSessionId !== hostSessionId || !message.user)
3689
+ return "accepted";
3690
+ if (lease.checkpoint?.answered || lease.pendingReply) {
3691
+ this.deactivate(hostSessionId);
3692
+ return "accepted";
3693
+ }
3694
+ if (lease.checkpoint)
3695
+ delete lease.checkpoint.advance;
3696
+ if (!lease.checkpoint && lease.inFlight !== "status")
3697
+ this.deactivate(hostSessionId);
3698
+ else {
3699
+ lease.messageId = messageId;
3700
+ lease.delivery = delivery;
3701
+ lease.pendingReply = true;
3702
+ if (!lease.inFlight)
3703
+ await this.onIdle(hostSessionId);
3609
3704
  }
3610
3705
  return "accepted";
3611
3706
  }
3612
3707
  compactionContext(hostSessionId) {
3613
- return this.#lease?.hostSessionId === hostSessionId ? "An in-memory /flow-auto continuation remains active. Follow authoritative compact Flow state after compaction without expanding the approved goal; stop for required user direction, a hard blocker, or confirmed closure." : null;
3708
+ if (this.#lease?.hostSessionId !== hostSessionId)
3709
+ return null;
3710
+ const context = [
3711
+ "An in-memory /flow-auto continuation remains active.",
3712
+ "Read compact Flow state first.",
3713
+ CONTINUATION_ROUTE,
3714
+ "Do not expand the approved goal; stop for required user direction, a hard blocker, or confirmed closure."
3715
+ ].join(" ");
3716
+ return `${context}
3717
+
3718
+ ${FLOW_MANAGER_KERNEL}`;
3719
+ }
3720
+ observeHostMessage(host, message) {
3721
+ const lease = this.#lease;
3722
+ if (lease?.hostSessionId !== host)
3723
+ return;
3724
+ if (message.role === "assistant" && message.parentID !== undefined) {
3725
+ lease.assistantParents.set(message.id, message.parentID);
3726
+ if (message.summary !== true) {
3727
+ lease.lastAssistantParent = message.parentID;
3728
+ if (lease.compaction)
3729
+ lease.compaction = null;
3730
+ } else if (lease.compaction) {
3731
+ if (message.parentID === lease.compaction.user && lease.messageId === lease.compaction.authority)
3732
+ lease.compaction.summary = message.id;
3733
+ else
3734
+ lease.compaction = null;
3735
+ }
3736
+ } else if (message.role === "user" && lease.compaction) {
3737
+ const compaction = lease.compaction;
3738
+ if (compaction.summary && (!compaction.successor || compaction.successor === message.id))
3739
+ compaction.successor = message.id;
3740
+ else
3741
+ lease.compaction = null;
3742
+ }
3743
+ }
3744
+ observeHostPart(host, part) {
3745
+ const lease = this.#lease;
3746
+ if (lease?.hostSessionId !== host || part.type !== "compaction" || part.auto !== true)
3747
+ return;
3748
+ lease.compaction = lease.lastAssistantParent && lease.lastAssistantParent === lease.messageId ? { authority: lease.lastAssistantParent, user: part.messageID } : null;
3749
+ }
3750
+ observeCompaction(host) {
3751
+ const lease = this.#lease;
3752
+ if (lease?.hostSessionId !== host || lease.messageId === null)
3753
+ return;
3754
+ const compaction = lease.compaction;
3755
+ lease.compaction = null;
3756
+ if (!compaction?.successor || lease.messageId !== compaction.authority)
3757
+ return void this.#rejectOrigin(lease, "compaction");
3758
+ lease.messageId = compaction.successor;
3759
+ }
3760
+ observeMutation(host, revision, created, assistantId, reviewerPending) {
3761
+ const lease = this.#lease;
3762
+ if (lease?.hostSessionId !== host || !lease.messageId)
3763
+ return;
3764
+ const origin = lease.assistantParents.get(assistantId);
3765
+ if (origin === undefined)
3766
+ return void this.#rejectOrigin(lease, "mutation");
3767
+ if (origin !== lease.messageId)
3768
+ return;
3769
+ const baseline = lease.baseline;
3770
+ if (baseline && baseline.sessionId === undefined && created)
3771
+ lease.baseline = { ...baseline, sessionId: created };
3772
+ const point = lease.checkpoint;
3773
+ if (!point)
3774
+ return;
3775
+ if (revision > point.revision)
3776
+ point.advance = revision + Number(reviewerPending);
3614
3777
  }
3615
3778
  timingSnapshot() {
3616
3779
  const timing = this.#timing;
@@ -3628,9 +3791,14 @@ class AutoDriveCoordinator {
3628
3791
  }
3629
3792
  async onIdle(hostSessionId) {
3630
3793
  const lease = this.#lease;
3631
- if (!lease || lease.hostSessionId !== hostSessionId || lease.inFlight)
3794
+ if (!lease || lease.hostSessionId !== hostSessionId)
3632
3795
  return;
3633
- lease.inFlight = "status";
3796
+ if (lease.inFlight) {
3797
+ if (lease.inFlight !== "reply-status" || !lease.pendingReply || lease.checkpoint?.advance !== undefined)
3798
+ lease.idlePending = true;
3799
+ return;
3800
+ }
3801
+ lease.inFlight = lease.pendingReply ? "reply-status" : "status";
3634
3802
  try {
3635
3803
  const projection = await this.#read(lease);
3636
3804
  if (!projection)
@@ -3638,88 +3806,72 @@ class AutoDriveCoordinator {
3638
3806
  if (this.#lease !== lease)
3639
3807
  return;
3640
3808
  const baseline = lease.baseline;
3641
- if (!baseline) {
3642
- this.#stop(lease);
3643
- return;
3644
- }
3645
- if (projection.status === "idle" || projection.nextAction === null) {
3646
- this.deactivate(hostSessionId);
3647
- return;
3648
- }
3649
- if (lease.flowSessionId && projection.sessionId !== lease.flowSessionId) {
3650
- this.#stop(lease, "Flow auto-drive stopped: Flow session changed.");
3651
- return;
3652
- }
3653
- lease.flowSessionId = projection.sessionId ?? lease.flowSessionId;
3809
+ if (!baseline)
3810
+ return this.#stop(lease);
3811
+ const anchored = lease.checkpoint !== null;
3812
+ if (projection.status === "idle" || projection.nextAction === null)
3813
+ return void this.deactivate(hostSessionId);
3814
+ if (projection.sessionId !== baseline.sessionId)
3815
+ return this.#stop(lease, "Flow auto-drive stopped: unowned session.");
3816
+ const checkpoint = lease.checkpoint;
3817
+ const boundary = isCheckpoint(projection);
3818
+ const advance = checkpoint?.advance;
3819
+ const mutationAdvanced = advance !== undefined && projection.revision === advance && isMechanical(projection);
3654
3820
  if (lease.pendingReply) {
3655
- const delivery = lease.pendingReply;
3656
- lease.pendingReply = null;
3657
- if (!lease.checkpoint && !isCheckpoint(projection)) {
3658
- this.deactivate(hostSessionId);
3659
- return;
3660
- }
3661
- lease.checkpoint = {
3662
- revision: projection.revision,
3663
- answered: true
3664
- };
3665
- lease.delivery = delivery;
3821
+ lease.pendingReply = false;
3822
+ if (boundary && (!checkpoint || projection.revision > checkpoint.revision))
3823
+ return void this.#waitAt(lease, projection.revision);
3824
+ if (!checkpoint || !boundary && !mutationAdvanced)
3825
+ return void this.deactivate(hostSessionId);
3826
+ checkpoint.answered = true;
3666
3827
  this.#setTiming("active");
3667
3828
  return;
3668
3829
  }
3669
- if (lease.checkpoint && !lease.checkpoint.answered)
3670
- return;
3671
- if (lease.checkpoint) {
3672
- if (projection.revision <= lease.checkpoint.revision) {
3673
- this.deactivate(hostSessionId);
3674
- return;
3675
- }
3676
- if (isCheckpoint(projection)) {
3677
- lease.checkpoint = { revision: projection.revision, answered: false };
3678
- this.#setTiming("waiting-for-user");
3679
- return;
3680
- }
3681
- if (!isMechanical(projection)) {
3682
- this.deactivate(hostSessionId);
3683
- return;
3684
- }
3685
- lease.checkpoint = null;
3686
- }
3687
- if (!isMechanical(projection)) {
3688
- if (isCheckpoint(projection)) {
3689
- lease.checkpoint = { revision: projection.revision, answered: false };
3690
- lease.lastPromptedRevision = null;
3691
- this.#setTiming("waiting-for-user");
3692
- } else {
3693
- this.deactivate(hostSessionId);
3694
- }
3695
- return;
3830
+ if (boundary) {
3831
+ if (checkpoint && projection.revision < checkpoint.revision)
3832
+ return void this.deactivate(hostSessionId);
3833
+ return void this.#waitAt(lease, projection.revision);
3696
3834
  }
3697
- if (baseline.sessionId ? projection.revision <= baseline.revision : projection.sessionId === undefined) {
3698
- this.#stop(lease, "Flow auto-drive stopped: initiating turn made no progress.");
3699
- return;
3835
+ if (checkpoint) {
3836
+ if (projection.revision <= checkpoint.revision || !mutationAdvanced)
3837
+ return void this.deactivate(hostSessionId);
3838
+ lease.checkpoint = null;
3700
3839
  }
3840
+ if (!isMechanical(projection))
3841
+ return void this.deactivate(hostSessionId);
3701
3842
  if (lease.lastPromptedRevision === projection.revision) {
3702
3843
  this.#setTiming("paused");
3703
- this.#warn(`Flow auto-drive paused after revision ${projection.revision} made no lifecycle progress.`);
3704
- return;
3705
- }
3706
- if (!lease.delivery) {
3707
- this.#stop(lease, "Flow auto-drive stopped: originating delivery was unavailable.");
3708
- return;
3844
+ return this.#warn(`Flow auto-drive paused after revision ${projection.revision} made no lifecycle progress.`);
3709
3845
  }
3846
+ if (baseline.sessionId ? projection.revision <= baseline.revision || !anchored && !isPendingReviewer(baseline) : projection.sessionId === undefined)
3847
+ return this.#stop(lease, "Flow auto-drive stopped: no progress.");
3848
+ if (!lease.delivery)
3849
+ return this.#stop(lease, "Flow auto-drive stopped: no delivery.");
3710
3850
  lease.lastPromptedRevision = projection.revision;
3851
+ lease.messageId = null;
3711
3852
  this.#setTiming("active");
3712
3853
  lease.inFlight = "prompt";
3713
3854
  try {
3714
- await this.#options.prompt(hostSessionId, `Continue the same user-authorized /flow-auto lifecycle from compact revision ${projection.revision}. Call flow_status with the compact view first, then follow ${projection.nextAction} without expanding the approved goal.`, lease.delivery, { [FLOW_AUTO_METADATA_KEY]: lease.token });
3855
+ const continuation = [
3856
+ `Continue the same user-authorized /flow-auto lifecycle from compact revision ${projection.revision}.`,
3857
+ "Call flow_status with the compact view first.",
3858
+ CONTINUATION_ROUTE,
3859
+ `Then follow ${projection.nextAction} without expanding the approved goal.`
3860
+ ].join(" ");
3861
+ await this.#options.prompt(hostSessionId, `${continuation}
3862
+
3863
+ ${FLOW_MANAGER_KERNEL}`, lease.delivery, { [FLOW_AUTO_METADATA_KEY]: lease.token });
3715
3864
  } catch (error) {
3716
- if (this.#lease === lease) {
3717
- this.#stop(lease, `Flow auto-drive stopped because continuation could not be enqueued: ${error instanceof Error ? error.message : String(error)}`);
3718
- }
3865
+ this.#stop(lease, `Flow auto prompt failed: ${String(error)}`);
3719
3866
  }
3720
3867
  } finally {
3721
- if (this.#lease === lease)
3868
+ if (this.#lease === lease) {
3869
+ const rerun = lease.idlePending;
3870
+ lease.idlePending = false;
3722
3871
  lease.inFlight = null;
3872
+ if (rerun)
3873
+ await this.onIdle(hostSessionId);
3874
+ }
3723
3875
  }
3724
3876
  }
3725
3877
  }
@@ -4294,11 +4446,29 @@ class ValidationCaptureCoordinator {
4294
4446
  }
4295
4447
 
4296
4448
  // src/platform/opencode/plugin.ts
4449
+ var MUTATION = /^flow_(?:plan_save|plan_approve|run_start|review_start|feature_complete|feature_reset|session_close)$/;
4450
+ var AUTO_STOPPED = "Flow auto stopped.";
4297
4451
  function isFlowCommand(command) {
4298
4452
  return Object.hasOwn(FLOW_CORE_COMMANDS, command);
4299
4453
  }
4300
- function commandPrompt(command, args) {
4301
- return FLOW_CORE_COMMANDS[command].template.replaceAll("$ARGUMENTS", () => args);
4454
+ function acceptedMutation(tool2, output) {
4455
+ if (!MUTATION.test(tool2))
4456
+ return null;
4457
+ try {
4458
+ const response = JSON.parse(output);
4459
+ const data = response.workflowData;
4460
+ const closeAccepted = tool2 === "flow_session_close" && response.status === "error" && data?.closeState?.durableAccepted === true;
4461
+ const revision2 = data?.projection?.revision;
4462
+ if (data?.operation?.replayed !== false || response.status !== "ok" && !closeAccepted || typeof revision2 !== "number" || !Number.isSafeInteger(revision2))
4463
+ return null;
4464
+ const sessionId = data.projection?.sessionId;
4465
+ return {
4466
+ revision: revision2,
4467
+ sessionId: typeof sessionId === "string" ? sessionId : undefined
4468
+ };
4469
+ } catch {
4470
+ return null;
4471
+ }
4302
4472
  }
4303
4473
  function textPart(text2, synthetic = false, metadata) {
4304
4474
  return {
@@ -4308,69 +4478,71 @@ function textPart(text2, synthetic = false, metadata) {
4308
4478
  ...metadata ? { metadata } : {}
4309
4479
  };
4310
4480
  }
4311
- function rewriteManagerCommand(command, args, output) {
4312
- if (output.parts.some((part) => part.type === "subtask")) {
4313
- throw new Error("Flow manager commands cannot contain subtask parts.");
4481
+ function rewriteCommand(command, args, output) {
4482
+ const config = FLOW_CORE_COMMANDS[command];
4483
+ const promptArgs = config.subtask ? args : "the preceding non-synthetic Flow request";
4484
+ const prompt = config.template.split("$ARGUMENTS").join(promptArgs);
4485
+ if (!config.subtask) {
4486
+ if (output.parts.some((part) => part.type === "subtask"))
4487
+ throw new Error("Flow manager commands cannot contain subtask parts.");
4488
+ const preserved = output.parts.filter((part) => part.type !== "text");
4489
+ output.parts.splice(0, output.parts.length, textPart(args.trim() ? `Flow ${command}: ${args}` : `Flow ${command}`), textPart(prompt, true), ...preserved);
4490
+ return;
4314
4491
  }
4315
- const preserved = output.parts.filter((part) => part.type !== "text");
4316
- output.parts.splice(0, output.parts.length, textPart(args.trim() ? `Flow ${command}: ${args.trim()}` : `Flow ${command}`), textPart(commandPrompt(command, args), true), ...preserved);
4317
- }
4318
- function rewriteReviewerCommand(command, args, output) {
4319
- if (output.parts.length !== 1 || output.parts[0]?.type !== "subtask") {
4492
+ if (output.parts.length !== 1 || output.parts[0]?.type !== "subtask")
4320
4493
  throw new Error(`/${command} requires exactly one reviewer subtask.`);
4321
- }
4322
4494
  const subtask = output.parts[0];
4323
- const config = FLOW_CORE_COMMANDS[command];
4324
- if (!config.subtask || subtask.agent !== config.agent) {
4325
- throw new Error(`/${command} must dispatch to '${config.subtask ? config.agent : "none"}'.`);
4326
- }
4327
- if (subtask.command?.replace(/^\/+/, "") !== command) {
4495
+ if (subtask.agent !== config.agent)
4496
+ throw new Error(`/${command} must dispatch to '${config.agent}'.`);
4497
+ if (subtask.command?.replace(/^\/+/, "") !== command)
4328
4498
  throw new Error(`/${command} subtask identity did not match.`);
4329
- }
4330
- subtask.prompt = commandPrompt(command, args);
4499
+ subtask.prompt = prompt;
4331
4500
  }
4332
4501
  function createCommandHook(assertOperational, autoDrive) {
4333
4502
  return async (input, output) => {
4334
4503
  const command = input.command.replace(/^\/+/, "");
4335
4504
  if (!isFlowCommand(command))
4336
4505
  return;
4337
- assertOperational(`execute /${command}`);
4338
- if (FLOW_CORE_COMMANDS[command].subtask) {
4339
- rewriteReviewerCommand(command, input.arguments, output);
4340
- } else {
4341
- rewriteManagerCommand(command, input.arguments, output);
4506
+ const action = input.arguments.trim();
4507
+ if (command === "flow-auto" && /^(?:stop|cancel)$/i.test(action)) {
4508
+ const confirmed = output.parts.some((part) => part.type === "text" && part.text === AUTO_STOPPED);
4509
+ const response = autoDrive.deactivate(input.sessionID) || confirmed ? AUTO_STOPPED : "No Flow auto lease was active in this OpenCode session.";
4510
+ output.parts[0] = textPart(response);
4511
+ output.parts.length = 1;
4512
+ return;
4342
4513
  }
4343
- if (command !== "flow-auto") {
4514
+ assertOperational(`execute /${command}`);
4515
+ rewriteCommand(command, input.arguments, output);
4516
+ if (command !== "flow-auto")
4517
+ return void autoDrive.deactivate(input.sessionID);
4518
+ const metadata = await autoDrive.activate(input.sessionID);
4519
+ const instruction = output.parts.find((part) => part.type === "text" && part.synthetic === true);
4520
+ if (!instruction) {
4344
4521
  autoDrive.deactivate(input.sessionID);
4345
- } else {
4346
- const metadata = await autoDrive.activate(input.sessionID);
4347
- const instruction = output.parts.find((part) => part.type === "text" && part.synthetic === true);
4348
- if (!instruction) {
4349
- autoDrive.deactivate(input.sessionID);
4350
- throw new Error("/flow-auto is missing its synthetic instruction.");
4351
- }
4352
- instruction.metadata = {
4353
- ...instruction.metadata ?? {},
4354
- ...metadata
4355
- };
4522
+ throw new Error("/flow-auto is missing its synthetic instruction.");
4356
4523
  }
4524
+ instruction.metadata = { ...instruction.metadata, ...metadata };
4357
4525
  };
4358
4526
  }
4359
- function guardTools(tools, runtimeGuard) {
4527
+ function guardTools(tools, runtimeGuard, autoDrive) {
4360
4528
  return Object.fromEntries(Object.entries(tools).map(([name, definition]) => [
4361
4529
  name,
4362
4530
  {
4363
4531
  ...definition,
4364
4532
  execute: async (...args) => {
4365
4533
  const status = runtimeGuard.query();
4366
- if (!status.operational) {
4534
+ if (!status.operational)
4367
4535
  return JSON.stringify({
4368
4536
  status: "error",
4369
4537
  summary: status.message,
4370
4538
  workflowData: { runtimeGuard: status }
4371
4539
  });
4372
- }
4373
- return definition.execute(...args);
4540
+ const output = await definition.execute(...args);
4541
+ const mutation = acceptedMutation(name, String(output));
4542
+ const context = args[1];
4543
+ if (mutation)
4544
+ autoDrive.observeMutation(context.sessionID, mutation.revision, name === "flow_plan_save" && mutation.revision === 1 ? mutation.sessionId : undefined, context.messageID, name === "flow_review_start");
4545
+ return output;
4374
4546
  }
4375
4547
  }
4376
4548
  ]));
@@ -4385,7 +4557,8 @@ var FlowPlugin = async (ctx) => {
4385
4557
  instanceId: createFlowPluginInstanceId()
4386
4558
  });
4387
4559
  const initial = runtimeGuard.query();
4388
- log(initial.operational ? "info" : "error", `Flow ${version}: ${initial.message}`);
4560
+ const level = initial.operational ? "info" : "error";
4561
+ log(level, `Flow ${version}: ${initial.message}`);
4389
4562
  const workspace = ctx.worktree ?? ctx.directory;
4390
4563
  const autoDrive = new AutoDriveCoordinator({
4391
4564
  readProjection: async () => {
@@ -4411,14 +4584,7 @@ var FlowPlugin = async (ctx) => {
4411
4584
  body: {
4412
4585
  agent: delivery.agent,
4413
4586
  model: delivery.model,
4414
- parts: [
4415
- {
4416
- type: "text",
4417
- text: prompt,
4418
- synthetic: true,
4419
- metadata: { ...metadata }
4420
- }
4421
- ]
4587
+ parts: [textPart(prompt, true, metadata)]
4422
4588
  },
4423
4589
  throwOnError: true
4424
4590
  });
@@ -4437,16 +4603,15 @@ var FlowPlugin = async (ctx) => {
4437
4603
  config: createConfigHook(ctx, {
4438
4604
  assertOperational: (action) => runtimeGuard.assertOperational(action)
4439
4605
  }),
4440
- tool: guardTools(tools, runtimeGuard),
4606
+ tool: guardTools(tools, runtimeGuard, autoDrive),
4441
4607
  "command.execute.before": createCommandHook((action) => runtimeGuard.assertOperational(action), autoDrive),
4442
4608
  "chat.message": async (input, output) => {
4443
4609
  const observed = await autoDrive.observeMessage(input.sessionID, {
4444
4610
  agent: output.message.agent,
4445
4611
  model: output.message.model
4446
- }, output.parts);
4447
- if (observed === "stale-continuation") {
4612
+ }, output.parts, output.message.id);
4613
+ if (observed === "stale-continuation")
4448
4614
  throw new Error("Discarded a stale Flow auto continuation.");
4449
- }
4450
4615
  },
4451
4616
  "experimental.session.compacting": async (input, output) => {
4452
4617
  const context = autoDrive.compactionContext(input.sessionID);
@@ -4455,32 +4620,28 @@ var FlowPlugin = async (ctx) => {
4455
4620
  },
4456
4621
  event: async (input) => {
4457
4622
  const event = input.event;
4623
+ if (event.type === "message.updated")
4624
+ return autoDrive.observeHostMessage(event.properties.info.sessionID, event.properties.info);
4625
+ if (event.type === "message.part.updated")
4626
+ return autoDrive.observeHostPart(event.properties.part.sessionID, event.properties.part);
4458
4627
  if (event.type === "session.deleted" || event.type === "session.error") {
4459
4628
  const sessionID2 = event.type === "session.deleted" ? event.properties.info.id : event.properties.sessionID;
4460
- if (sessionID2) {
4461
- validation.cancel(sessionID2);
4462
- autoDrive.deactivate(sessionID2);
4463
- } else {
4464
- autoDrive.clear();
4465
- }
4466
- return;
4629
+ if (!sessionID2)
4630
+ return autoDrive.clear();
4631
+ validation.cancel(sessionID2);
4632
+ return void autoDrive.deactivate(sessionID2);
4467
4633
  }
4468
- if (event.type !== "session.idle" && event.type !== "session.compacted") {
4634
+ if (event.type !== "session.idle" && event.type !== "session.compacted")
4469
4635
  return;
4470
- }
4471
4636
  const sessionID = event.properties?.sessionID;
4472
4637
  validation.cancel(sessionID);
4473
- if (event.type === "session.idle") {
4474
- if (runtimeGuard.query().operational) {
4475
- await autoDrive.onIdle(sessionID);
4476
- } else {
4477
- autoDrive.deactivate(sessionID);
4478
- }
4479
- }
4480
- },
4481
- "tool.execute.before": async (input, output) => {
4482
- validation.observeToolBefore(input, output);
4638
+ if (event.type === "session.compacted")
4639
+ return autoDrive.observeCompaction(sessionID);
4640
+ if (runtimeGuard.query().operational)
4641
+ return autoDrive.onIdle(sessionID);
4642
+ autoDrive.deactivate(sessionID);
4483
4643
  },
4644
+ "tool.execute.before": async (input, output) => validation.observeToolBefore(input, output),
4484
4645
  "tool.execute.after": async (input, output) => {
4485
4646
  try {
4486
4647
  await validation.observeToolAfter(input, output);
@@ -4503,4 +4664,4 @@ export {
4503
4664
  plugin_default as default
4504
4665
  };
4505
4666
 
4506
- //# debugId=F1C0E773B6734DA364756E2164756E21
4667
+ //# debugId=454D701E6238F47E64756E2164756E21