plat 128

← Pulse platform issue index

PLAT-128 — guidance tests assert content for two docs that were deliberately deleted 8 days before

Coordination	Value
Assigned agent	unassigned
Ticket state	implemented (5 of 24 baseline failures); remainder is unrelated wording drift, deferred
Last synchronized	2026-08-17

• Priority: P3 — nothing product-facing was wrong; a test suite was asserting behavior for a doc the product no longer has.
• Owner: cmd/server/guidance/render_all_test.go, pkg/orchestrator/agents/workflow/step_based_workflow prompt tests

How it surfaced

Auditing tool-error logs for PLAT-125–127 kept surfacing a permanently-red 24-failure baseline in cmd/server, guidance, virtual-tools, and step_based_workflow. A red baseline is exactly what let real regressions hide earlier in this diagnosis — worth clearing rather than living with.

Two unrelated root causes, not one

A test/production drift, unrelated to any deleted doc. TestReviewPlanPromptPrefersCoherentAgenticSteps and TestReadOnlyExecutionPromptCannotRecommendMutationOrRawSQLite build their own hand-written templateVars map and call ExecuteTemplate/the prompt processor directly. The first was missing "DBGuidance", a key every real caller (WorkflowPlanReviewAgent.Execute) sets before rendering the same template — a plain test bug. The second asserted the read-only execution prompt must never contain the substring mutate_workflow_db anywhere, which became impossible once the guidance was made more explicit: "This session is read-only: do not call mutate_workflow_db." (BuildManagedWorkflowDBGuidance, prompt_sections.go:79) necessarily names the tool it forbids. Naming the forbidden tool is correct, explicit guidance — the same shape as this session's record_run_concern refusal fix (PLAT-123's identity-gap redirect) and the tools_unavailable message itself. The test's own name says the real intent: cannot recommend mutation, not cannot mention it.

A deliberate architecture change the tests never caught up with. git log --diff-filter=D -- '*review-improve-log*' '*review-code*' finds 0174b6aff "simplify Pulse workflow reviews", 2026-08-08 — 9 days before this ticket. That commit deleted templates/review/review-code.md, templates/review/bug-review.md, and (per an earlier commit in the same lineage) the review-improve-log / review-improve-log-migration / review-improve-log-skeleton trio, and rewrote goal-advisor.md from 538 lines down to a fraction of that. render_all_test.go still called renderFromRegistry("review-improve-log", ...) and renderFromRegistry("review-code", ...) directly — kinds absent from both guidance.go's registry and the filesystem — across 8 test functions and roughly 90 assertion lines describing builder/improve.html's exact HTML contract: no_terminal_packet, retry_due, approved_awaiting_evidence, a skeleton markup spec, an archive migration contract. improve.html itself is confirmed fully retired — Pulse's own DB-backed findings (record_pulse_finding, record_pulse_result) and the SQLite Pulse popup replaced it.

review-code (main.py-vs-step-description drift checking) has no live caller anywhere in the codebase today, and its own test assertion expected data-module="bug_review" — the other deleted kind's identity, not its own — evidence the two were never cleanly separated even before removal. Owner decision: leave it removed from the test suite; whether to design a replacement is a separate, deliberate task, not something to restore unexamined from an 8-day-stale file.

Fix

• execution_only_agent_test.go, maintenance_reviewer_prompts_test.go: add the missing DBGuidance key; change the read-only-prompt check from substring-absence to asserting the explicit prohibition is present and no recommending phrase ("Use mutate_workflow_db") is.
• render_all_test.go: removed every review-improve-log/-migration/ -skeleton/review-code assertion — whole functions where that was their entire content (TestReviewImproveLogMigrationIsExtracted), single map entries or list items elsewhere. TestPulseSpecialistsReturnStructuredPacketsAndParentOwnsHTML additionally dropped a "builder/improve.html" content requirement, confirmed absent from all 10 remaining live specialist templates by rendering each one — not a guess, and not specific to review-code.

Verification

Fail-before/pass-after (stash the 3 changed files, rerun, restore). Full 4-package regression run on current origin/main: 24 pre-existing failures before, 19 after, zero new — the 5 fixed here are exactly the ones this ticket addresses.

Not fixed here

• 19 failures remain, all single-phrase wording drift in templates that are still live and correctly registered (pulse-setup, improve-evaluation, goal-advisor, scheduler preflight-count tracking, and others) — a different, larger body of work needing individual judgment per assertion about whether the test or the content drifted. Not attempted here.
• Whether review-code (or a successor) should exist is an open product question, deliberately not decided by this ticket.