Document continuous execution prompt
Some checks are pending
M6 deployable RC / quick (push) Waiting to run
M6 deployable RC / chromium (push) Blocked by required conditions
M6 deployable RC / release (push) Blocked by required conditions

This commit is contained in:
mes123456
2026-08-23 05:54:08 -04:00
parent 9f43244982
commit c0d471da01

View File

@@ -11,42 +11,62 @@ message. The loop stops only at an explicit stop condition described below.
## Copyable Resume Prompt ## Copyable Resume Prompt
Use the following prompt when continuous execution is needed. It deliberately Use the following prompt when continuous execution is needed. It is the
keeps every task iteration bounded while allowing successful handoffs to proceed copyable, end-to-end command for bounded implementation, verification, handoff,
automatically: and (when authorized) repository push:
```text ```text
按 nextTask.md 接续执行。 按 nextTask.md 接续执行。
Treat this as an execution command, not a request for a plan or status report. Treat this as an execution command, not a request for a plan or status report.
Enter CONTINUOUS mode and repeat the following bounded iteration until a stop Enter CONTINUOUS mode and repeat one bounded task iteration at a time until a
condition is reached: stop condition is reached:
1. Read docs/EXECUTION_QUEUE.md, run print-task-context, then run 1. Read `docs/EXECUTION_QUEUE.md`; run `node tools/web/print-task-context.mjs`
check-task-context before editing. and `node tools/web/check-task-context.mjs` before editing.
2. Read only the current task card, parent manifest/status, and the focused 2. Read only the current task card, parent manifest/status, and the focused
files named by the task context. Do not open the full next-task-plan.json, files named by the task context. Never open the full
all status logs, or historical project plans. `next-task-plan.json`, all `docs/status` logs, or historical plans.
3. Complete exactly the current task: implement the scoped behavior, run the 3. Complete exactly the current task: implement the scoped behavior, generate
focused desktop/npm/direct checks, verify save/reopen and the negative case, the fixture, run the focused desktop Blender check, npm comparator, and
write reports/status/manifest/task-context with SHA-256 values, then run direct comparator, then verify save/reopen and the malformed/negative case.
governance checks and git diff --check. Chromium is the only permitted browser runtime.
4. Advance the queue only after every exit criterion and hash check passes, 4. Keep long output in `tests/golden/<task>/command.log`; report only exit
using repository generators for the task index/catalog/card. codes, short tails, artifact paths, and SHA-256 values.
5. After a successful handoff, discard the previous task transcript and begin 5. After every exit criterion passes, run
the next iteration from a fresh print-task-context package. Do not wait for `node tools/web/check-task-context.mjs --task <task-id> --write`, write the
another user message and do not carry prior task files into the next report/status/manifest/task-context handoff, and verify all artifact hashes.
iteration except through the new parent manifest and status summary. 6. Run every post-handoff gate:
Keep tool output and context bounded. On any failure or hash drift, keep the ```bash
current task `in_progress`/`blocked`, do not advance `nextTask`, and stop the node tools/web/check-blender-next-task-plan.mjs
loop. End only when a stop condition is reached, with a compact summary of all node tools/web/check-task-index.mjs
tasks completed in this run (task, state, nextTask, command exit codes, and npm --prefix web run test:task-context
artifact paths/hashes). npm --prefix web run test:context-governance
For stream-disconnect or remote-compact errors, stop the active request and node tools/web/check-context-governance.mjs
use the recovery procedure below in a new request. Do not paste the failed git diff --check
transcript, invoke another compact operation, or blindly repeat a state- ```
changing command.
7. Only after all gates pass, use repository generators for the task
index/catalog/card and advance `docs/EXECUTION_QUEUE.md` to the newly
completed manifest's `nextTask`. Do not manually edit task-index offsets,
catalog lines, parent evidence, or a guessed task number.
8. If repository saving has been authorized, commit the successful handoff and
push the current branch to its existing `origin`; use the credential helper
or an interactive prompt, never write credentials into files, URLs, logs, or
prompts. Verify `HEAD` and `origin/<branch>` match after the push.
9. After a successful handoff, discard the previous task transcript and begin
the next iteration from a fresh `print-task-context` package. Do not wait for
another user message and carry forward only the new parent manifest/status.
On any failed command, missing focused entry, unavailable environment, budget
overflow, hash drift, governance failure, cancellation, or inconsistent queue,
keep the task `in_progress`/`blocked`, do not advance `nextTask`, and stop with
the exact blocker and evidence paths. Do not fake completion or retry an
unknown state-changing command. For stream-disconnect, error-decoding, or
remote-compact failures, use the recovery procedure below in a new request.
End with a compact summary of completed tasks, states, nextTask, command exit
codes, artifact hashes, push status, and any remaining blocker.
``` ```
## Resume Command ## Resume Command