Gate settlement reply carries step outputs and passed-step feedback

jared commented

2026-08-23 17:47:49 +00:00

Owner

Context

Since the #477 in-turn gate redesign, the programmer/test-writer agents run os-sdlc-runner gate <dispatch_id> themselves, but the reply built at lib/os_sdlc/runner/gate_settlement.rb:64-67 contains only {type, dispatch_id, verdict, step}. The full gate output is already captured per step as GateStepOutcome rows (gate_verdict.rb:30) but surfaces only via the gate_results enrich on the NEXT dispatch — a SubagentStop-era design. An agent told only verdict: fail, step: test must re-run the suite manually to see why, doubling every iteration's cost (an 80s suite becomes 160s per round).

Tasks

GateVerdict collects its GateStepOutcome objects and returns them alongside the Verdict (or Verdict gains passed/failed step data).
GateSettlement#record includes in the reply hash: the failing step's captured output, the list of steps that passed, and an instruction line telling the agent to correct the failures and rerun os-sdlc-runner gate <dispatch_id>.
On full pass, the reply states all steps green and that the agent may finish.
Runner CLI gate command prints the enriched reply so the invoking agent sees it in-turn.
Tests cover fail-with-output, pass-with-prior-steps, and full-green reply shapes.

Acceptance criteria

A failing lint gate reply names the offenses (rubocop output) and the steps already green, without the agent running any test/lint command itself.
No schema change; StepResult capture path unchanged.

Origin

Trigger: #476 grilling session 2026-08-23, trace of the os-sdlc-runner gate path.
Improvised this session: none.
Chain: blind in-turn iteration ← reply built from Verdict only (gate_settlement.rb:64-67) ← DESIGN (feedback routed to the next dispatch via gate_results enrich, ADR-0133/0134 era SubagentStop model).
Root candidate: this ticket.

## Context Since the #477 in-turn gate redesign, the programmer/test-writer agents run `os-sdlc-runner gate <dispatch_id>` themselves, but the reply built at lib/os_sdlc/runner/gate_settlement.rb:64-67 contains only `{type, dispatch_id, verdict, step}`. The full gate output is already captured per step as GateStepOutcome rows (gate_verdict.rb:30) but surfaces only via the `gate_results` enrich on the NEXT dispatch — a SubagentStop-era design. An agent told only `verdict: fail, step: test` must re-run the suite manually to see why, doubling every iteration's cost (an 80s suite becomes 160s per round). ## Tasks - [ ] GateVerdict collects its GateStepOutcome objects and returns them alongside the Verdict (or Verdict gains passed/failed step data). - [ ] GateSettlement#record includes in the reply hash: the failing step's captured output, the list of steps that passed, and an instruction line telling the agent to correct the failures and rerun `os-sdlc-runner gate <dispatch_id>`. - [ ] On full pass, the reply states all steps green and that the agent may finish. - [ ] Runner CLI `gate` command prints the enriched reply so the invoking agent sees it in-turn. - [ ] Tests cover fail-with-output, pass-with-prior-steps, and full-green reply shapes. ## Acceptance criteria - [ ] A failing lint gate reply names the offenses (rubocop output) and the steps already green, without the agent running any test/lint command itself. - [ ] No schema change; StepResult capture path unchanged. ## Origin - Trigger: #476 grilling session 2026-08-23, trace of the `os-sdlc-runner gate` path. - Improvised this session: none. - Chain: blind in-turn iteration ← reply built from Verdict only (gate_settlement.rb:64-67) ← DESIGN (feedback routed to the next dispatch via `gate_results` enrich, ADR-0133/0134 era SubagentStop model). - Root candidate: this ticket.

jared added the

plugin/os-sdlc

label

2026-08-23 17:47:49 +00:00

jared referenced this issue

2026-08-23 17:48:46 +00:00

Decide target resolution for the session-less os-sdlc CLI (red-assert/green-assert TargetFlag) #476

jared commented

2026-08-23 17:53:37 +00:00

Author

Owner

Work starting on branch worktree-ticket-490 via /os-sdlc:implement pipeline.

jared referenced this issue

2026-08-23 18:22:29 +00:00

Fix brief composer's false 'no tracker configured' in subagent-start briefs #494

jared commented

2026-08-23 18:22:55 +00:00

Author

Owner

Resolution

Done: Gate settlement reply now carries step outputs and passed-step feedback: GateVerdict collects its GateStepOutcome rows (outcomes accessor); GateSettlement builds a GateReply value object whose reply hash adds passed_steps, the failing step's captured output, and a message (rerun instruction naming the dispatch id on fail; 'all gates green -- you may finish' on pass). CLI gate prints the enriched reply via the existing emit path. No schema change; StepResult capture path untouched.

Evidence: Branch worktree-ticket-490. Files: plugins/os-sdlc/lib/os_sdlc/runner/gate_verdict.rb, gate_settlement.rb, new tests/runner/gate_settlement_test.rb (full-green, fail-with-output, fail-with-prior-passed-steps shapes, incl. rubocop offense text in lint-fail reply output). Pipeline gates green: 1085 runs, 0 failures; rubocop no offenses. Reviewer (opus) verdict: pass.

Follow-ups: #494 captured (brief composer false 'no tracker configured', skeptic CREATE). Lint-output assertion done inline this session (skeptic DROP-do-inline). Dropped: private_constant GateReply, collapse record_verdict/record, fail-open failing_output — reviewer-rated non-blocking, no provenance beyond style.

## Resolution **Done:** Gate settlement reply now carries step outputs and passed-step feedback: GateVerdict collects its GateStepOutcome rows (outcomes accessor); GateSettlement builds a GateReply value object whose reply hash adds passed_steps, the failing step's captured output, and a message (rerun instruction naming the dispatch id on fail; 'all gates green -- you may finish' on pass). CLI gate prints the enriched reply via the existing emit path. No schema change; StepResult capture path untouched. **Evidence:** Branch worktree-ticket-490. Files: plugins/os-sdlc/lib/os_sdlc/runner/gate_verdict.rb, gate_settlement.rb, new tests/runner/gate_settlement_test.rb (full-green, fail-with-output, fail-with-prior-passed-steps shapes, incl. rubocop offense text in lint-fail reply output). Pipeline gates green: 1085 runs, 0 failures; rubocop no offenses. Reviewer (opus) verdict: pass. **Follow-ups:** #494 captured (brief composer false 'no tracker configured', skeptic CREATE). Lint-output assertion done inline this session (skeptic DROP-do-inline). Dropped: private_constant GateReply, collapse record_verdict/record, fail-open failing_output — reviewer-rated non-blocking, no provenance beyond style.

jared closed this issue

2026-08-23 18:22:56 +00:00

jared referenced this issue from a commit

2026-08-23 18:31:46 +00:00

feat(os-sdlc): gate settlement reply carries step outputs and passed-step feedback (#490)

jared referenced this issue from a commit

2026-08-23 18:31:46 +00:00

Merge worktree-ticket-490: gate settlement reply enrichment (#490)

jared referenced this issue from a commit

2026-08-23 18:31:46 +00:00

fix(os-sdlc): drop stale 'Single job' prefix from code-probe description

jared referenced this issue from a commit