Skip to content

t1452: enforce pulse dispatch launch validation and CLI guardrails #4193

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
marcusquinn merged 1 commit into from
Mar 12, 2026
Merged

t1452: enforce pulse dispatch launch validation and CLI guardrails #4193

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 18 additions & 0 deletions .agents/scripts/commands/pulse.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -558,6 +558,22 @@ sleep 2

If a dispatch attempt exits immediately with provider/auth failure (for example `Token refresh failed`, `authentication`, `401`, `403`, `400` in startup logs), do not wait for next cycle. Re-dispatch in the same cycle via `headless-runtime-helper.sh run` with an explicit alternate model/provider and continue filling remaining slots.

**Launch validation is mandatory (t1452):** after each dispatch, verify the worker actually started and did not fall through to CLI usage output.

```bash
# 1) Process exists (session-key must map to a live run)
pgrep -fal "issue-<number>|Issue #<number>:" >/dev/null
Copy link

@gemini-code-assist gemini-code-assist bot Mar 12, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The pgrep command can be simplified. The -a (show full command line) and -l (show process name) flags are used for display purposes, but since the output is redirected to /dev/null, they are unnecessary. Using only the -f flag is sufficient to check for the process's existence via its exit code and is more idiomatic.

Suggested change
pgrep -fal "issue-<number>|Issue #<number>:" >/dev/null
pgrep -f "issue-<number>|Issue #<number>:" >/dev/null


# 2) Startup log is not CLI help/usage output
LOG_PATH="/tmp/pulse-<slug>-<number>.log"
if [[ -f "$LOG_PATH" ]] && rg -q '^opencode run \[message\.\.]|^run opencode with a message|^Options:' "$LOG_PATH"; then
echo "Invalid worker launch for #<number> (CLI usage output)"
# Relaunch immediately via helper (never leave this for next pulse)
fi
```

If validation fails, re-dispatch immediately via `headless-runtime-helper.sh run`, add a short issue comment noting the failed launch and correction, and continue filling slots.
Comment on lines +563 to +575
Copy link
Contributor

@coderabbitai coderabbitai bot Mar 12, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Potential issue | 🟠 Major

Use the helper’s real launch contract here, not an invented /tmp log file.

headless-runtime-helper.sh run does not write startup output to /tmp/pulse-<slug>-<number>.log; it captures to a temp file internally, validates activity, and deletes that file before returning. As written, this “mandatory” check will usually no-op on the log check and miss the exact usage-output false starts from #4192. Please move this to the deterministic helper/check path the issue calls for, or have the helper emit a launch-validation result that pulse can consume directly.

🤖 Prompt for AI Agents 
Verify each finding against the current code and only fix it if needed.

In @.agents/scripts/commands/pulse.md around lines 563 - 575, The pulse script's
current startup validation wrongly inspects a /tmp log (LOG_PATH) that
headless-runtime-helper.sh does not leave behind; replace this ad-hoc file check
with the helper's deterministic launch-validation contract: call
headless-runtime-helper.sh run (or its dedicated validate/launch subcommand) and
consume its explicit exit status or emitted validation result/JSON rather than
grepping LOG_PATH, then on failure immediately re-dispatch via
headless-runtime-helper.sh run and post the short issue comment; update
references in pulse to use the helper's validation output (instead of
LOG_PATH/rg) and adjust the relaunch logic to rely on the helper's
success/failure signal.


9. **Fill-to-cap post-condition (t1449):** before ending the pulse cycle, compare active workers vs `MAX_WORKERS`. If below cap and runnable scoped issues/PR work exists in any repo class, continue dispatching until cap is reached or no runnable candidates remain. Do not leave slots idle because of class reservations when one class is PR-capped or empty.

### Candidate discovery baseline (t1443 + t1448)
Expand Down Expand Up @@ -588,7 +604,9 @@ If you dispatch an unassigned issue without `auto-dispatch`/`status:available`,
- Background with `&`, sleep 2 between dispatches
- The helper alternates the default headless providers/models (`anthropic/claude-sonnet-4-6`, `openai/gpt-5.3-codex`), persists session IDs per provider + session key, honors provider backoff, and rejects `opencode/*` gateway models (no Zen fallback for headless runs)
- Do NOT add `--model` for first attempts — let the helper choose the alternating default. **Exception:** when escalating after 2+ failed attempts on the same issue, pass `--model anthropic/claude-opus-4-6` to the helper (see "Model escalation after repeated failures" above).
- If you must run raw `opencode run` for diagnosis, use only documented flags from `opencode run --help` and NEVER pass unsupported options (for example `--max-iterations`); unsupported flags cause usage-output false starts that burn worker slots.
- If helper-selected launch fails at startup with auth/provider errors, immediately retry with explicit alternate provider in the same cycle (for example `--model openai/gpt-5.3-codex` after anthropic auth failure) and log the fallback in an issue comment.
- After every dispatch, run launch validation (live process + no CLI usage output in startup log) before counting the slot as filled.
- Use `--dir <path>` from repos.json
- Route non-code tasks with `--agent`: SEO, Content, Marketing, Business, Research (see AGENTS.md "Agent Routing")
- If a dispatched worker later looks stalled, `worker-watchdog.sh` now inspects the recent OpenCode transcript tail before killing it, includes that diagnostic evidence in the retry trail, and gives provider-wait evidence one extra timeout window before re-queueing the issue.
Expand Down
Loading