fix: re-trace system prompt on agent switch

[VJ-yadav]

Summary

• The step === 1 guard in prompt.ts only traced the first agent's system prompt because step increments continuously and is never reset on agent switches
• Added a tracedAgent variable to track the last-traced agent name, so each agent's system prompt is logged when it first becomes active
• Fixed the inaccurate comment that claimed agent switches reset step to 0

Fixes #291

Test Plan

• Verify typecheck passes on prompt.ts (pre-existing test file failures are unrelated)
• Start a session, switch agents mid-conversation, confirm both agents' system prompts appear in trace output
• Confirm single-agent sessions still trace the system prompt exactly once

Checklist

• Change is minimal and focused on the reported issue
• No functional behavior change — tracing/observability only
• Comment updated to accurately describe the logic

Summary by CodeRabbit

• Refactor
  • Improved session tracing so system-prompt entries are emitted on the initial step and whenever execution switches to a different agent. This reduces redundant trace noise, clarifies agent transitions in logs, and ensures each agent’s system prompt appears once per switch for cleaner, more readable session traces.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: c220a8e6-80c3-4a17-b4d9-a4676300359a

📥 Commits

Reviewing files that changed from the base of the PR and between 3c7fe0c and 01c48e6.

📒 Files selected for processing (1)

• packages/opencode/src/session/prompt.ts

🚧 Files skipped from review as they are similar to previous changes (1)

• packages/opencode/src/session/prompt.ts

---

📝 Walkthrough

Walkthrough

System-prompt tracing in SessionPrompt.loop now tracks the last-traced agent name and logs the "system-prompt" span when either the loop is at step 1 or the active agent differs from the tracked agent, then updates the tracked agent name so each agent's system prompt is traced once upon activation.

Changes

Cohort / File(s)	Summary
Agent Tracing Logic packages/opencode/src/session/prompt.ts	Add loop-local tracedAgent state; change Tracer guard from step === 1 to `step === 1

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Suggested labels

contributor

Poem

🐰 I hopped through loops with nimble feet,

I watched each agent's name I’d greet.
When new friends came, I left a trace,
No prompt omitted in my race.
Hop, log, repeat — a tidy feat.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 inconclusive)

Check name	Status	Explanation	Resolution
Description check	❓ Inconclusive	The PR description is mostly complete with Summary, Test Plan, and Checklist sections, though it lacks the required 'PINEAPPLE' marker at the top as specified in the template.	Add 'PINEAPPLE' at the very top of the PR description before any other content, as required by the repository template.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: re-tracing the system prompt when agents switch, which is the primary objective of the PR.
Linked Issues check	✅ Passed	The PR successfully implements the primary fix from issue #291 [#291]: introducing a tracedAgent variable to track agent switches and re-emit tracing spans when agent identity changes, satisfying the core requirement.
Out of Scope Changes check	✅ Passed	All changes are tightly scoped to the reported issue #291: tracking agent identity changes to re-trace system prompts and updating an inaccurate comment, with no unrelated modifications.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[dev-punia-altimate]

❌ Tests — Failures Detected

TypeScript — 15 failure(s)

• connection_refused [2.73ms]
• timeout [2.41ms]
• permission_denied [2.67ms]
• parse_error [2.35ms]
• oom [2.49ms]
• network_error [2.33ms]
• auth_failure [2.50ms]
• rate_limit [2.59ms]
• internal_error [2.63ms]
• empty_error [0.25ms]
• connection_refused [0.16ms]
• timeout [0.10ms]
• permission_denied [0.07ms]
• parse_error [0.07ms]
• oom [0.07ms]

cc @VJ-yadav
_{Tested at 3b691b5a | Run log | Powered by QA Autopilot}

[anandgupta42]

Review — Ready after rebase

Scope: Fixes #291 — step === 1 guard in prompt.ts only traced the first agent's system prompt because step increments continuously and is never reset on agent switches (despite the comment claiming otherwise).

Analysis

• Minimal, targeted fix: 6 additions, 5 deletions in one file (session/prompt.ts)
• Introduces tracedAgent to track the last-traced agent name; logs each agent's system prompt when it first becomes active
• Fixes the misleading comment that claimed step resets to 0 on agent switches
• Pure observability/tracing fix — no functional runtime behavior change
• Correctness matches the issue's root-cause analysis exactly

CI status (commit 3b691b5a)

• ⚠️ Only check-standards, check-compliance, GitGuardian ran — missing TypeScript, Marker Guard, anti-slop
• Same reason as fix: add pagination support for recap trace list #596: PR is ~1 week behind main, CI workflow matrix has since expanded. Rebase will retrigger full CI
• ⚠️ session/prompt.ts IS an upstream-shared file — after rebase, verify Marker Guard passes (altimate_change markers needed around the tracedAgent variable + the restructured if-block)
• centralized-test-results — same batch-wide flakiness as every other open PR

Recommendation
✅ Ready for merge after rebase + Marker Guard verification. The fix itself is correct and minimal. The only real risk here is the upstream-shared file needing markers — confirm by running bun run script/upstream/analyze.ts --markers --base main --strict locally post-rebase.

cc @anandgupta42 — rebase, verify marker guard, then ready for /release.

[anandgupta42]

Consensus code-review applied — no changes needed

Ran a 4-participant consensus review (Claude + GPT 5.4 + Gemini 3.1 Pro + self-review).

Verdict: APPROVE — both GPT 5.4 and Gemini 3.1 Pro agree the 6-line fix is correct and ready to merge.

The tracedAgent guard at packages/opencode/src/session/prompt.ts correctly traces each agent's system prompt exactly once per distinct contiguous agent, which is the stated contract from issue #291. Resumption safety verified — tracedAgent="" default preserves re-capture-on-fresh-loop behavior.

Non-blocking suggestions (NOT applied, kept as optional follow-ups):

• Regression test for multi-iteration loop with an agent switch (GPT, MAJOR / Gemini, MINOR) — would require stubbing Tracer.active?.logSpan and driving a mock plan→builder switch
• A→B→A edge case test to lock in current re-trace-on-return behavior (GPT, MINOR)
• Simplify the redundant step === 1 check since tracedAgent="" and agent.name is always non-empty (Gemini, NIT)

The first suggestion is worthwhile but requires session-loop test infrastructure that doesn't exist yet; tracked as a follow-up. The fix is small enough that shipping without new tests is acceptable.

All CI green. Ready for merge as-is.