CON-216: stop campaign_assistant replaying stale prior instructions by grsmv · Pull Request #115 · ogen-app/ogen

grsmv · 2026-08-15T11:37:08Z

Fixes CON-216 — "post_assistant seems stuck in previous user message instructions in case of an internal error."

Problem

A user asked a read-only question ("tell me about assets attached to this campaign"), but the cheap Haiku planner:

Spuriously fired generatePosts (a write action nobody asked for) alongside the correct askCampaignAssets, and
Reused a date (2026-07-28) that appears only in earlier conversation history, not the current message.

resolveWindow rejected the past date with a Go error, which aborts the whole genkit Generate turn → surfaced to the user as a raw red model call failed: … is in the past and discarded the good asset answer.

What was already fixed

Layer 2 (the crash) was resolved by CON-215: generatePosts/draftPost now fail soft on a past date / non-target platform / unknown phase (a zero-post warning, not a turn-aborting error), leaving the turn conversational. The red 502 in the screenshot can no longer recur.
CON-213 added max-turns graceful recovery + the one-heavy-action-per-turn latch.

This PR addresses Layer 1 — the root-cause mis-routing named in the ticket title — plus a slot-leak the mis-route exposed.

Changes

Prompt discipline (prompts/campaign_assistant.tmpl): act only on the current message; treat history as background, not a queue to replay. Questions are read-only and must never trigger a write/generation tool; resolve dates only from the current request against today — never reuse a date from an earlier message.
Late heavy-action reservation (toolGeneratePosts): reserve the turn's single heavy slot only after the user-correctable validations pass, mirroring the CON-207 draftPost refactor. A mis-routed/invalid generatePosts that fails soft no longer burns the slot for a legitimate heavy action in the same turn. ensureCampaignAssetUse (which persists UseAssets to the DB) also moved past the reservation, so a skipped/soft-failed turn never writes state for work that won't run.
Regression test (tools_test.go): TestGeneratePosts_HeavySlotOnlyBurnedOnSuccess — a soft-failed generatePosts leaves the slot claimable; a successful one claims it (second heavy reservation turned away).

Why no server-side date guard

Dates are model-resolved and history is legitimately needed (draftPost sources its material from it), and the crash path is already handled. Prompt discipline + the reservation reorder is the proportionate fix.

Testing

go build, go vet, and go test -race ./src/genkit/flows/campaign_assistant/ all pass (existing CON-213/214/215 tests still green).
Template parses with both system/context blocks intact.

Summary by CodeRabbit

Bug Fixes
- Improved the assistant’s handling of conversation context, ensuring actions are based only on the current request.
- Invalid or incomplete post-generation requests now fail gracefully without affecting campaign state.
- Heavy actions are reserved only after request validation succeeds, preventing valid follow-up actions from being blocked.

The Haiku planner could re-issue a write tool (e.g. generatePosts) for a read-only question, reusing a date pulled from conversation history rather than the current message — the "stuck in previous user message" bug. Two targeted fixes: - Prompt discipline (campaign_assistant.tmpl): act only on the current message; treat history as background, not a queue to replay. Questions are read-only and never trigger a write/generation tool; resolve dates only from the current request against today, never reuse a date from history. - Late heavy-action reservation (toolGeneratePosts): reserve the turn's single heavy slot only after the user-correctable validations pass, mirroring draftPost (CON-207). A mis-routed/invalid generatePosts that fails soft (CON-215) no longer burns the slot for a legitimate heavy action in the same turn. ensureCampaignAssetUse also moved past the reservation so a skipped turn never flips UseAssets in the DB. Layer 2 (the past-date abort surfacing as a raw "model call failed") was already resolved by CON-215's soft-fail; this addresses the root-cause mis-routing. Adds a regression test that the heavy slot is burned only on a successful generation.

linear-code · 2026-08-15T11:37:11Z

CON-216

coderabbitai · 2026-08-15T11:37:25Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: a83b914a-b27d-43a8-822e-640cc2f286b9

📥 Commits

Reviewing files that changed from the base of the PR and between 9d96909 and a8e492a.

📒 Files selected for processing (3)

src/genkit/flows/campaign_assistant/prompts/campaign_assistant.tmpl
src/genkit/flows/campaign_assistant/tools.go
src/genkit/flows/campaign_assistant/tools_test.go

Walkthrough

The campaign assistant now uses only current-message instructions for write actions and dates. generatePosts validates requests before consuming the per-turn heavy-action slot or changing asset state.

Changes

Campaign assistant behavior

Layer / File(s)	Summary
Current-message request scoping `src/genkit/flows/campaign_assistant/prompts/campaign_assistant.tmpl`	The prompt treats conversation history as background, handles questions read-only, requires explicit current-message write requests, and resolves dates from the current request.
Validated heavy-action reservation `src/genkit/flows/campaign_assistant/tools.go`, `src/genkit/flows/campaign_assistant/tools_test.go`	`generatePosts` validates inputs before reserving the heavy-action slot. Asset-use initialization occurs after reservation. Tests cover soft failure, successful reservation, and rejection of a second reservation.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Merge Risk: ⚪ Minimal · up to a8e49

The change limits stale-instruction replay and prevents invalid generation attempts from consuming the turn's heavy-action slot; no actionable merge-blocking risk remains after normal checks and review.

Sequence Diagram(s)

sequenceDiagram
  participant UserMessage
  participant CampaignAssistant
  participant generatePosts
  participant HeavyActionSlot
  participant GenerationFlow
  UserMessage->>CampaignAssistant: current request
  CampaignAssistant->>generatePosts: explicit generation request
  generatePosts->>generatePosts: validate request
  generatePosts->>HeavyActionSlot: reserve after valid checks
  generatePosts->>GenerationFlow: generate posts
  GenerationFlow-->>CampaignAssistant: generated posts

Possibly related PRs

ogen-app/ogen#111: Introduced the per-turn heavy-action latch that this change refines.
ogen-app/ogen#113: Modified generatePosts validation and soft-failure behavior.
ogen-app/ogen#114: Modified the Campaign Assistant prompt and tool handling refined here.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the primary change: preventing campaign_assistant from replaying stale instructions from conversation history.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

Create stacked PR
Commit on current branch

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch fix/con-216-campaign-assistant-stale-instruction-misroute

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

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

grsmv · 2026-08-15T15:00:19Z

@CodeRabbit review

coderabbitai · 2026-08-15T15:00:24Z

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

grsmv deployed to testing August 15, 2026 11:37 — with GitHub Actions Active

grsmv added the to test label Aug 15, 2026

grsmv deployed to testing August 15, 2026 11:41 — with GitHub Actions Active

grsmv deployed to testing August 15, 2026 11:43 — with GitHub Actions Active

ogen-app deleted a comment from coderabbitai Bot Aug 15, 2026

grsmv merged commit cf2c146 into main Aug 15, 2026
6 checks passed

grsmv deleted the fix/con-216-campaign-assistant-stale-instruction-misroute branch August 15, 2026 21:30

Navigation Menu

Uh oh!

CON-216: stop campaign_assistant replaying stale prior instructions - #115

CON-216: stop campaign_assistant replaying stale prior instructions#115
grsmv merged 1 commit into
mainfrom
fix/con-216-campaign-assistant-stale-instruction-misroute

grsmv commented Aug 15, 2026 •

edited by coderabbitai Bot

Loading

Uh oh!

linear-code Bot commented Aug 15, 2026

Uh oh!

coderabbitai Bot commented Aug 15, 2026 •

edited

Loading

❌ Failed checks (1 warning)

Uh oh!

grsmv commented Aug 15, 2026

Uh oh!

coderabbitai Bot commented Aug 15, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

grsmv commented Aug 15, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

What was already fixed

Changes

Why no server-side date guard

Testing

Summary by CodeRabbit

Uh oh!

linear-code Bot commented Aug 15, 2026

Uh oh!

coderabbitai Bot commented Aug 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Sequence Diagram(s)

❌ Failed checks (1 warning)

Uh oh!

grsmv commented Aug 15, 2026

Uh oh!

coderabbitai Bot commented Aug 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

grsmv commented Aug 15, 2026 •

edited by coderabbitai Bot

Loading

coderabbitai Bot commented Aug 15, 2026 •

edited

Loading

coderabbitai Bot commented Aug 15, 2026 •

edited

Loading