@staff0rd/assist 0.1160.0 → 0.1161.0

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.
Files changed (127) hide show
  1. package/agent/advice/ask-check.md +1 -1
  2. package/agent/advice/ask-question.md +1 -1
  3. package/agent/commands/add-rule.md +2 -2
  4. package/agent/commands/bug.md +1 -1
  5. package/agent/commands/design.md +1 -0
  6. package/agent/commands/draft.md +1 -1
  7. package/agent/commands/github.md +2 -0
  8. package/agent/commands/pr.md +2 -0
  9. package/agent/commands/refine.md +1 -1
  10. package/agent/commands/review-config.md +2 -0
  11. package/agent/commands/review-high-level.md +1 -1
  12. package/dist/.build-files +107 -106
  13. package/dist/{addActivity-TNE2FRPB.js → addActivity-FI4P6OTC.js} +3 -3
  14. package/dist/{addPhase-V4Z636NN.js → addPhase-TAU5TNIH.js} +8 -6
  15. package/dist/{advise-RTJV7NU2.js → advise-P6UNZDSJ.js} +5 -4
  16. package/dist/ask-F6E63RQQ.js +50 -0
  17. package/dist/{askCheck-3JCSRRAQ.js → askCheck-Y5ZASXYZ.js} +13 -12
  18. package/dist/{askDesign-Q4HBBLJD.js → askDesign-TZKXMGMV.js} +8 -7
  19. package/dist/{askQuestion-QXVDXZZY.js → askQuestion-PDKTLFTV.js} +6 -4
  20. package/dist/{branch-LSX2L22U.js → branch-6T6R4B3N.js} +4 -4
  21. package/dist/built-version +1 -1
  22. package/dist/{chunk-C4247LMI.js → chunk-26N4PV7Z.js} +1 -1
  23. package/dist/chunk-3VDV7VHX.js +6 -0
  24. package/dist/{chunk-OTYBOGWB.js → chunk-4X2D3DCV.js} +1 -1
  25. package/dist/{chunk-4X2UUOFH.js → chunk-627WPDA6.js} +1 -1
  26. package/dist/{chunk-CKJ4SDVJ.js → chunk-6GUFWQ3N.js} +1 -1
  27. package/dist/{chunk-TESRDGY7.js → chunk-6JWOFHSW.js} +1 -1
  28. package/dist/{chunk-C6AKA3GY.js → chunk-74UYKNXC.js} +2 -2
  29. package/dist/{chunk-7I2Q27D6.js → chunk-77ZKOKNF.js} +7 -7
  30. package/dist/{chunk-Q4CX4OJJ.js → chunk-7KTG4VNA.js} +1 -1
  31. package/dist/{chunk-I3XPML6O.js → chunk-7NRDYDQO.js} +6 -6
  32. package/dist/{chunk-6ZECLGYV.js → chunk-7VK2KF6L.js} +8 -5
  33. package/dist/{chunk-VEMNENJ2.js → chunk-AGBHU2F2.js} +1 -1
  34. package/dist/{chunk-WQF27PEJ.js → chunk-ASM6OJRK.js} +1 -1
  35. package/dist/{chunk-S55TRXFP.js → chunk-BEZEB6OF.js} +2 -2
  36. package/dist/{chunk-L4PVKBXS.js → chunk-BP6SFWR6.js} +1 -1
  37. package/dist/{chunk-5THMDW42.js → chunk-CMQE2SXC.js} +1 -1
  38. package/dist/{chunk-MSPCQ5L7.js → chunk-CVSVONYT.js} +1 -1
  39. package/dist/{chunk-K5K7LATN.js → chunk-DJ5UUPZR.js} +1 -1
  40. package/dist/{chunk-5AJWUYA3.js → chunk-EIA6AM33.js} +1 -1
  41. package/dist/{chunk-MOQ2IEPI.js → chunk-EXGBNBBJ.js} +1 -1
  42. package/dist/{chunk-EOALZG4Y.js → chunk-F5AIU5Q3.js} +1 -1
  43. package/dist/{chunk-7MIAK3SD.js → chunk-FJQSHCU3.js} +2 -2
  44. package/dist/{chunk-OLFUSXO3.js → chunk-GZT3NLC6.js} +1 -1
  45. package/dist/{chunk-L3RJER2C.js → chunk-HEP6K2DT.js} +1 -1
  46. package/dist/{chunk-D2337HWD.js → chunk-HN5RRUW2.js} +1 -1
  47. package/dist/{chunk-SUO2246Y.js → chunk-I6B6W6WT.js} +1 -1
  48. package/dist/{chunk-O7X46AAU.js → chunk-ILRYEIH3.js} +1 -1
  49. package/dist/{chunk-MG4IJB47.js → chunk-J265JKFW.js} +3 -3
  50. package/dist/{chunk-25PQYD5B.js → chunk-JZ2TTDJ5.js} +9 -3
  51. package/dist/{chunk-SQKNRK6D.js → chunk-LPBBNHBJ.js} +1 -1
  52. package/dist/{chunk-LO4K3OUQ.js → chunk-LVDWS4SI.js} +1 -1
  53. package/dist/{chunk-JR3LZHBZ.js → chunk-LX6IOH5J.js} +1 -1
  54. package/dist/{chunk-GMXZUWBU.js → chunk-NA5TFFHL.js} +1 -1
  55. package/dist/{chunk-K4TQJQWA.js → chunk-OQ6AYXHB.js} +1 -1
  56. package/dist/{chunk-K5MYLLFM.js → chunk-OX4K4FAY.js} +2 -2
  57. package/dist/{chunk-JVJOXBO2.js → chunk-QEFHPIMP.js} +3 -3
  58. package/dist/{chunk-54QTHCXV.js → chunk-QRTB7PVR.js} +7 -1
  59. package/dist/{chunk-VWVLMTH2.js → chunk-QS2S63AF.js} +4 -0
  60. package/dist/{chunk-VYOVPMBH.js → chunk-QUWL7IBN.js} +1 -1
  61. package/dist/{chunk-ERAGZGT5.js → chunk-RMGPNDEN.js} +2 -0
  62. package/dist/chunk-TQ5KKZ6G.js +8 -0
  63. package/dist/{chunk-K4T24UGW.js → chunk-UE5PIMY2.js} +1 -1
  64. package/dist/{chunk-LYQWAGLE.js → chunk-UM3CASEM.js} +1 -1
  65. package/dist/{chunk-YHZUTP5K.js → chunk-WQ3ABZ4N.js} +1 -1
  66. package/dist/{chunk-QSUS3RLF.js → chunk-XTD6227R.js} +1 -1
  67. package/dist/{chunk-XDQFHWGO.js → chunk-YZSTCQUE.js} +4 -4
  68. package/dist/{chunk-C5YM3YO4.js → chunk-Z4LG3KDK.js} +35 -1
  69. package/dist/{cliHook-XFH6QLH7.js → cliHook-E5HX4RV7.js} +7 -6
  70. package/dist/{cliHookCheck-QXSRKVUJ.js → cliHookCheck-KQDTKJBA.js} +3 -3
  71. package/dist/{codexHook-KGR76TSJ.js → codexHook-EMML4X3I.js} +9 -8
  72. package/dist/commands/extension/extension/updates.json +2 -2
  73. package/dist/commands/sessions/web/bundle.js +1 -1
  74. package/dist/{comment-2T2VC3NL.js → comment-AOB7KRXL.js} +7 -4
  75. package/dist/{commentIssue-KOQ5PSTG.js → commentIssue-6G3N2E4B.js} +6 -3
  76. package/dist/{commit-DXTHQVIK.js → commit-KOTH6G6L.js} +4 -4
  77. package/dist/{createIssue-EE5W4LLB.js → createIssue-IQJL3QO4.js} +5 -2
  78. package/dist/{doctorNodes-HBQW5L3M.js → doctorNodes-5EW6QQ5D.js} +3 -3
  79. package/dist/{done-MQBZZSGP.js → done-G57I4IXY.js} +5 -5
  80. package/dist/{drainDaemon-T2X2H2Z5.js → drainDaemon-7V6RMDRT.js} +1 -1
  81. package/dist/{edit-OBXC5Z6J.js → edit-XLS74GEX.js} +5 -2
  82. package/dist/{editHook-PIDVOHE3.js → editHook-HDXB5IRZ.js} +7 -6
  83. package/dist/{editIssue-CR4SFOIS.js → editIssue-53DIWJIJ.js} +7 -4
  84. package/dist/{editIssueComment-SVTPJ6CU.js → editIssueComment-HCEYRKWT.js} +6 -3
  85. package/dist/{fixConflict-2DB66U4A.js → fixConflict-YPNQ3Z5O.js} +5 -5
  86. package/dist/{importBacklog-FOGUOAIJ.js → importBacklog-65MOC2KY.js} +2 -1
  87. package/dist/index.js +71 -71
  88. package/dist/{launchMode-4TWOI76X.js → launchMode-3PLZA4EV.js} +14 -14
  89. package/dist/{list-QFKV7MEU.js → list-EXCFXUUV.js} +6 -6
  90. package/dist/{merge-MVMNYPO2.js → merge-5FURAXPB.js} +3 -2
  91. package/dist/{next-TJGY3KDH.js → next-MELQDUE4.js} +13 -13
  92. package/dist/{piHook-7QDAYTIX.js → piHook-SXJBRASG.js} +7 -6
  93. package/dist/{postSlackMessage-MH6BMLGC.js → postSlackMessage-KXS4R7Y6.js} +6 -3
  94. package/dist/{propose-JF27BF42.js → propose-37TNXC2B.js} +11 -9
  95. package/dist/{raise-5YMRZJLP.js → raise-KEPF3WOY.js} +5 -2
  96. package/dist/readBodyArgument-WHNTOBUE.js +8 -0
  97. package/dist/{readTime-23CA6JVM.js → readTime-CWEUMPXR.js} +3 -2
  98. package/dist/{refine-DPA5BYQI.js → refine-VY42GT5E.js} +15 -15
  99. package/dist/{reply-RF5XWST2.js → reply-6MCVVNZK.js} +6 -3
  100. package/dist/{restartNode-I3C4XEWU.js → restartNode-VNJAREDR.js} +5 -5
  101. package/dist/{review-GXVALUWY.js → review-S5JIPTTJ.js} +11 -9
  102. package/dist/{reviewPrComments-CICNPWVX.js → reviewPrComments-PLRIHXWC.js} +5 -5
  103. package/dist/{rewindPhase-DDTS4N6S.js → rewindPhase-MKAAJZ3B.js} +9 -9
  104. package/dist/{rulesAdd-2PHHDPJ3.js → rulesAdd-T62MTCFM.js} +6 -3
  105. package/dist/{run-6AHR53L7.js → run-4VCFD6GC.js} +11 -11
  106. package/dist/{run-BDQCSTCF.js → run-IH2SPTOS.js} +11 -11
  107. package/dist/{runDaemon-FE3XZEUT.js → runDaemon-4WJQINVB.js} +71 -33
  108. package/dist/{runExtract-43B4PDWU.js → runExtract-4IIID534.js} +5 -2
  109. package/dist/{runRefine-ZNMBCAAN.js → runRefine-SM6EHRUB.js} +15 -15
  110. package/dist/{scan-NKKHHGL2.js → scan-Q2LQ6C2H.js} +3 -3
  111. package/dist/{schemaConfirm-6NZ3JBDT.js → schemaConfirm-DVNHG5CE.js} +3 -3
  112. package/dist/{schemaUnlock-N53NXNUF.js → schemaUnlock-AQAEVGUD.js} +3 -3
  113. package/dist/{search-X52HVMA3.js → search-PFHCJUFA.js} +6 -6
  114. package/dist/{show-N2KOJIBG.js → show-EQCAKDSF.js} +4 -4
  115. package/dist/{show-PR4OSVLY.js → show-NHSD46FE.js} +7 -6
  116. package/dist/{signExtension-WUVAF67T.js → signExtension-VIFT2IDU.js} +1 -1
  117. package/dist/{slack-4YK2TTH5.js → slack-SMNDERHS.js} +6 -3
  118. package/dist/{updateNode-PZMQ2BSE.js → updateNode-GUT5LN56.js} +5 -5
  119. package/dist/{updatePlan-IOSLPULZ.js → updatePlan-V57DKNB5.js} +11 -9
  120. package/dist/{web-H4GKNDJW.js → web-AQVFE5HY.js} +17 -17
  121. package/dist/{web-T44LOGKB.js → web-ZJ3S6UPP.js} +17 -17
  122. package/dist/{wontfix-62MHC5KB.js → wontfix-ASUD2OQM.js} +6 -3
  123. package/package.json +1 -1
  124. package/dist/ask-7JXPMU2X.js +0 -82
  125. package/dist/chunk-6MNX5ZMQ.js +0 -6
  126. package/dist/readBodyArgument-NZG3GJMF.js +0 -7
  127. package/dist/{chunk-7M657DAU.js → chunk-AFGBWFOC.js} +3 -3
@@ -3,4 +3,4 @@ title: Asking for a manual check
3
3
  when: askQuestion
4
4
  ---
5
5
 
6
- When you need the user to test something by hand, don't write the check up in chat. Write it as JSON (`{"title": ..., "steps": [...], "before": [...], "changed": ...}`: `title` is the expected outcome in at most 60 characters, `steps` are 1-4 steps of at most 90 characters each, optional `before` lists at most 4 short setup or restart chips, and optional `changed` is one sentence of at most 160 characters) to a scratch file and run `assist ask-check --json - < <file>` as a background task, doing no other work until it returns. It opens in the session's preview pane and blocks until the user answers Works or Doesn't work, printing the result and any note they added. Dismissing exits non-zero, meaning the user gave no result; don't assume it passed. Raise one check per thing to test, and keep any chat summary to a few lines. If it exits non-zero before opening (disabled, or the session isn't live), ask in chat instead. While a check is pending, a pasted question quoting a pending manual check is only a query: answer it in chat and keep waiting on the pending `ask-check`; don't re-run, supersede or stop it.
6
+ When you need the user to test something by hand, don't write the check up in chat. Write it as JSON (`{"title": ..., "steps": [...], "before": [...], "changed": ...}`: `title` is the expected outcome in at most 60 characters, `steps` are 1-4 steps of at most 90 characters each, optional `before` lists at most 4 short setup or restart chips, and optional `changed` is one sentence of at most 160 characters) to a scratch file and run `assist ask-check --json - < <file>` {{previewWait}}. It opens in the session's preview pane for the user to answer Works or Doesn't work, printing the result and any note they added. Dismissing exits non-zero, meaning the user gave no result; don't assume it passed. Raise one check per thing to test, and keep any chat summary to a few lines. If it exits non-zero before opening (disabled, or the session isn't live), ask in chat instead. While a check is pending, a pasted question quoting a pending manual check is only a query: answer it in chat and leave the pending `ask-check` open; don't re-run, supersede or stop it.
@@ -3,4 +3,4 @@ title: Asking questions
3
3
  when: askQuestion
4
4
  ---
5
5
 
6
- Don't use the built-in AskUserQuestion tool; the user answers from the session's web preview pane, which can't drive the terminal prompt. Write the same AskUserQuestion input as JSON (`{"questions": [...]}`, 1-4 questions each with `question` (at most 30 words), optional `context` (at most 40 words, shown above the question to say why you're asking), `header`, `multiSelect` and 2-4 `options` of `label`, `description` and optional `preview`) to a scratch file and run `assist ask-question --json - < <file>` as a background task, doing no other work until it returns. It opens as a form in the preview pane and blocks until answered. Answering exits 0 and prints the same answer text AskUserQuestion returns; dismissing exits non-zero, meaning the user declined to answer. To revise a still-pending question, re-run it directly; the earlier task then exits 0 as superseded — ignore its result.
6
+ Don't use the built-in AskUserQuestion tool; the user answers from the session's web preview pane, which can't drive the terminal prompt. Write the same AskUserQuestion input as JSON (`{"questions": [...]}`, 1-4 questions each with `question` (at most 30 words), optional `context` (at most 40 words, shown above the question to say why you're asking), `header`, `multiSelect` and 2-4 `options` of `label`, `description` and optional `preview`) to a scratch file and run `assist ask-question --json - < <file>` {{previewWait}}. It opens as a form in the preview pane. Answering exits 0 and prints the same answer text AskUserQuestion returns; dismissing exits non-zero, meaning the user declined to answer. {{previewRevise}}
@@ -59,10 +59,10 @@ It appends the rule to `advice.rules` in `user (db)` with no R-code. For `user (
59
59
 
60
60
  Use single quotes around the rule text to avoid shell escaping issues.
61
61
 
62
- Run `assist rules add` as a background task and do no other work until it returns. In a web session it opens the rule, its title and where it will land in the preview pane, and writes nothing until the user decides:
62
+ Run `assist rules add` as a background task and do no other work until it returns. In a web session it opens the rule, its title and where it will land in the preview pane, and writes nothing until the user decides. In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run `assist rules add` in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
63
63
 
64
64
  - **Approved** — it exits 0 having written the rule. Show the command output — it names the allocated code (or, for an `advice.rules` rule, the title) and where it went — then tell the user in one line which rule was added and where.
65
- - **Rejected** — it exits non-zero having written nothing, printing the reason and every inline comment. Revise the wording, title or scope to address each one, then re-run `assist rules add` with the revision as a background task again. Do not move on to Step 5 until a version is approved.
65
+ - **Rejected** — it exits non-zero having written nothing, printing the reason and every inline comment. Revise the wording, title or scope to address each one, then re-run `assist rules add` with the revision again, the same way. Do not move on to Step 5 until a version is approved.
66
66
  - **Superseded** — an earlier run whose preview a re-run replaced exits 0 saying so; ignore its result.
67
67
 
68
68
  Outside a web session it writes immediately and exits 0.
@@ -91,7 +91,7 @@ Omit `phases` when you are not authoring a plan. Each phase needs at least one t
91
91
 
92
92
  The payload is strict JSON — an unknown key is an error. `\n` inside the `description` string is a JSON escape and becomes a real newline, which is what the markdown rendering needs.
93
93
 
94
- Always run `propose` **as a background task**, and do no other work until it returns. In a web session it blocks on the preview pane until the user decides, which can take far longer than the default command timeout, and the pending preview dies with the process — a killed `propose` abandons the item. To revise the report while its preview is still pending, re-run `propose` with the revised payload directly — do not stop the earlier task. The pane swaps to the new version, and the earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without creating anything: ignore it.
94
+ Always run `propose` **as a background task**, and do no other work until it returns. In a web session it blocks on the preview pane until the user decides, which can take far longer than the default command timeout, and the pending preview dies with the process — a killed `propose` abandons the item. To revise the report while its preview is still pending, re-run `propose` with the revised payload directly — do not stop the earlier task. The pane swaps to the new version, and the earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without creating anything: ignore it. In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run `propose` in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
95
95
 
96
96
  `propose` decides where the bug report is reviewed and prints what to do next. Follow that instruction; never inspect the environment to work it out yourself:
97
97
 
@@ -50,6 +50,7 @@ assist ask-design --url '<artifact-url>' --title '<design title>' --body - < <su
50
50
  - **Approved** (exit 0) — fold every printed comment into the design, republishing to the same URL, then move on.
51
51
  - **Rejected** (non-zero, printing the reason and comments) — revise the design to address each one, republish to the same artifact URL, and re-run `ask-design` for the next round. Repeat until approved.
52
52
  - To revise while a round is still pending, re-run `ask-design` directly; the earlier task exits 0 as superseded — ignore its result.
53
+ - In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run `ask-design` in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
53
54
  - If it prints the design link and exits 0 without `Approved.` (not in an assist web session), ask the user in chat instead and iterate there — editing the same file and republishing to the same URL — until they say they are happy with it.
54
55
  - Without an `Artifact` tool there is no artifact URL, and `ask-design` refuses with the reason design review is unavailable on this harness. Review in chat instead: give the user the HTML file's absolute path and iterate on the same file until they say they are happy with it, then skip the link and attach steps below and summarise the design in the item's description.
55
56
 
@@ -121,7 +121,7 @@ JSON
121
121
 
122
122
  The payload is strict JSON — an unknown key is an error. `\n` inside the `description` string is a JSON escape and becomes a real newline, which is what the markdown rendering needs. Each phase needs at least one task; omit `manualChecks` for the phases that don't need any (most of them). A phase named `Review` is rejected: the runner appends that phase itself.
123
123
 
124
- Always run `propose` **as a background task**, and do no other work until it returns. In a web session it blocks on the preview pane until the user decides, which can take far longer than the default command timeout, and the pending preview dies with the process — a killed `propose` abandons the item. To revise the item while its preview is still pending, re-run `propose` with the revised payload directly — do not stop the earlier task. The pane swaps to the new version, and the earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without creating anything: ignore it.
124
+ Always run `propose` **as a background task**, and do no other work until it returns. In a web session it blocks on the preview pane until the user decides, which can take far longer than the default command timeout, and the pending preview dies with the process — a killed `propose` abandons the item. To revise the item while its preview is still pending, re-run `propose` with the revised payload directly — do not stop the earlier task. The pane swaps to the new version, and the earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without creating anything: ignore it. In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run `propose` in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
125
125
 
126
126
  `propose` decides where the item is reviewed and prints what to do next. Follow that instruction; never inspect the environment to work it out yourself:
127
127
 
@@ -64,6 +64,8 @@ Display the output so the user can see what happened. If the command exits non-z
64
64
 
65
65
  Any preview in this skill (`edit`, `create`, `comment`) can be revised while still pending by re-running the command directly — do not stop the earlier task. The earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without touching GitHub: ignore it.
66
66
 
67
+ In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run any of these previewing commands in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
68
+
67
69
  ## Action: create
68
70
 
69
71
  Create a new issue from a free-form request (e.g. `create bug against owner/repo, Sprint 3 milestone, Roadmap project: login fails on Safari`). Step 2 does not apply — there is no issue yet. Work out from the args:
@@ -6,6 +6,8 @@ Run `assist prs raise --help` and follow its guidance exactly to compose and rai
6
6
 
7
7
  ## After a successful raise
8
8
 
9
+ In a Claude SDK session (`ASSIST_SDK_AGENT=1`) a web-previewed raise returns at once: end your turn and do nothing else, and its outcome arrives as your next message.
10
+
9
11
  Only once the PR has actually been placed (skip this if the raise was rejected in the web preview or errored — there is nothing to watch), start watching CI in the background so you can keep working:
10
12
 
11
13
  - Run `gh pr checks --watch` as a background task. It resolves the current branch's PR on its own, so no arguments are needed. Do not block on it — continue with other work while it polls.
@@ -64,7 +64,7 @@ The payload is the whole plan, not a patch: every phase you want the item to end
64
64
 
65
65
  Do not chain `add-phase`, `update-phase`, `remove-phase` and `move-phase` calls to build up a change — that is one preview per command and leaves the item half-edited if the user rejects part way. Those commands remain for a human typing a quick one-liner at a terminal.
66
66
 
67
- In a web session this renders the whole change into the preview pane as one diff — each phase marked unchanged, added, edited or moved, removed phases listed separately, phases the item has already completed flagged, and any clamp of `currentPhase` into the shorter plan shown at the top — and blocks until the user decides. The review can take far longer than the default command timeout, so run `update-plan` **as a background task** and do no other work until it returns; the pending preview dies with the process. To revise the change while it is still pending, re-run `update-plan` with the revised payload directly — do not stop the earlier task. The pane swaps to the new diff, and the earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without writing anything: ignore it.
67
+ In a web session this renders the whole change into the preview pane as one diff — each phase marked unchanged, added, edited or moved, removed phases listed separately, phases the item has already completed flagged, and any clamp of `currentPhase` into the shorter plan shown at the top — and blocks until the user decides. The review can take far longer than the default command timeout, so run `update-plan` **as a background task** and do no other work until it returns; the pending preview dies with the process. To revise the change while it is still pending, re-run `update-plan` with the revised payload directly — do not stop the earlier task. The pane swaps to the new diff, and the earlier task exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` without writing anything: ignore it. In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run `update-plan` in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
68
68
 
69
69
  When it returns:
70
70
 
@@ -48,6 +48,8 @@ assist ask --title 'High-level review config' --body - < <scratch file>
48
48
  - **Rejected (non-zero exit)** — write nothing. Address the reason and each comment, re-check any changed glob still matches files, revise the scratch file and re-run `assist ask`.
49
49
  - **Superseded (exit 0)** — you re-ran `assist ask` while this one was pending, which is how to revise it; do not stop the earlier task first. Ignore this result and wait for the newer one.
50
50
 
51
+ In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run it in the foreground instead: it returns at once, so end your turn and do nothing else — the command's output and exit code arrive as your next message, and a superseded run delivers nothing.
52
+
51
53
  Outside a web session `assist ask` only prints the markdown and exits 0 without a decision: ask the user in chat instead, and write only once they have answered.
52
54
 
53
55
  ## Step 4: Write the answers
@@ -11,7 +11,7 @@ assist review --high-level $ARGUMENTS
11
11
 
12
12
  If this is a re-invocation in a session that already ran it, re-run it anyway — never assume the checklist is still open, since the pane may have gone even while a previous task still looks alive. Do not stop a still-running previous `assist review --high-level` task: launch it afresh, and the new checklist replaces the old one in place. The earlier task then exits 0 printing `Superseded by a newer preview — no decision; ignore this result.` — ignore it.
13
13
 
14
- Run it as a background task. In an assist web session it opens the checklist in the web UI preview pane and returns once the user approves or requests changes; otherwise it prints the checklist and exits straight away. Do not review the PR yourself or post anything to GitHub.
14
+ Run it as a background task. In an assist web session it opens the checklist in the web UI preview pane and returns once the user approves or requests changes; otherwise it prints the checklist and exits straight away. Do not review the PR yourself or post anything to GitHub. In a Claude SDK session (`ASSIST_SDK_AGENT=1`), run it in the foreground instead: it returns at once, and the review's output and exit code arrive as your next message once it closes; a superseded run delivers nothing.
15
15
 
16
16
  Once it is launched, say in one line that the high-level review has started and end your turn straight away. Do not say a pane is open — you only learn which case applies from the task's output. Do not wait on the task — no polling its output, `sleep`, `Monitor` or reading its output file — and the "do no other work until it returns" rule for sign-off commands does not apply here. Input the user sends while you are mid-turn only queues behind it, so staying in the turn blocks every question until the review closes.
17
17