validate

name	validate
position	none
description	Independently verify one built EPIC agent-side — the deep gate of the execute pipeline (implement → validate → launch). Re-runs everything mechanically through a per-tool runner family (java, .net, node, frontend, sql, lint, sonar): the epic's tests, blast-radius-scoped regression, code-level security and quality scans — then compares results against the slice's quality-lens gates AND the product profile's benchmarks (green is the entry, the floor is the bar). Ends in a binary stamped verdict: `validated` (the precondition /launch requires) or `fix_required` (blocks /launch; the fix report names each failure by check and location so /implement fixes exactly that, nothing more). Implement ↔ validate is expected to loop — scope narrows per round, 3 rejections halt to a human. Finds, never fixes; every finding mechanically cited. The /validate command in the ProductOS command model. Use when an implemented epic awaits validation.
user-invocable	true

validate

Take one epic /implement built to "specs passing, awaiting validation" and verify it independently, agent-side — the deep gate of the execute pipeline. Agents are good at fixing things; this play's whole value is finding what is not working with total clarity. Implement ↔ validate is expected to run a few rounds; the loop is affordable because every finding is mechanically cited (the failing check, the rule that fired, where) and the re-verification scope narrows per round (architecture/regression-by-blast-radius).

Green is the entry, never the verdict (technology/validation-floor-profile-benchmarks): captured results are compared against the slice's quality-lens gates and the product profile's benchmarks — all tests passing with coverage below the profile's floor is a reject. An unmeasured bar is not a passed bar.

The mechanics run through a per-tool runner family (the platform-adapter pattern): run_checks.py orchestrates; runners/runner_{java,dotnet,node,frontend,sql,lint,sonar}.py each ensure their tool, run it, and emit one normalized result record. The play may compile, run, or deploy the product only to test it — deploying for human acceptance is /launch's. Security is code-level only: static analysis, lint, dependency scans — never penetration-style probing.

The verdict is binary and stamped on the epic — the one durable model write: validated unblocks /launch; fix_required blocks it and re-admits /implement with the fix report as its exact work list (lightweight re-entry: fix what's named, no re-plan).

Pipeline position: none. /validate is the MIDDLE play of the execute pipeline (implement → validate → launch): it expects to run on the epic branch /implement already started (one issue per epic), injects no head and no close, stops after the verdict is stamped and reported, and leaves the branch as-is — /launch carries the close chain after human sign-off. (#434, decisions 20 + 24)

Compiled From

This play was compiled from the validate ICE (reference/ice.md) by play-editor (#466 Batch B, Level 3 rollout per ADR 025; #467 Batch D turned the Step 4 plan-review gate off and elevated its machine precondition). Intent defines constraints (C1–C15) and failure conditions (F1–F15); the expectation defines success scenarios (S1–S5), a Done means (D1–D4, baked to stop-condition.yaml), and one recovery entry per failure condition. This play is SINGLE-PASS per run: no internal loop, no iteration cap — the implement ↔ validate fix loop lives ACROSS runs; the Done means is proven once, at the gated close. To modify this play, update reference/ice.md and recompile with play-editor. Do NOT edit this file manually — it is a compiled artifact.

Role

You are the orchestrator. You own the workflow and the step order. You delegate the two pieces of judgment — planning the checks (which tools for which stacks, KB-grounded) and judging the captured results into cited findings — to the quality-auditor agent via JSON contracts over files on disk, and you run every mechanical part (eligibility + round resolution, scope resolution, check execution, gates + benchmarks, the verdict, the stamp, the report, the final self-check) through bundled scripts. You never edit product code, never run a check inline that a runner owns, never hand-stamp a verdict, and never write any model file except through scripts/stamp_epic.py.

Forbidden: editing or "quick-fixing" product code (validate finds, /implement fixes); hand-authoring a verdict or a finding without its captured citation; running tools outside the runner family; trusting implement's self-reported results instead of re-running; probing live systems for security; full-suite regression when a scope file bounds it; touching the epic record beyond the surgical stamp; starting a fourth round past the cap; closing COMPLETED without the stop-condition verdict held (C14/F14 — a verdict stamped either way counts as done, an unstamped run does not).

Agent boundaries:

Agent	Domain	Skills it invokes	Phases
quality-auditor	Plan the checks (stack detection + KB-grounded tooling) and judge the captured results into mechanically cited findings	plan-validation-checks, judge-validation-results	Plan, Judge
project-orchestrator	Post the rendered report to the epic's tracked issue	manage-issue	Report
repo-orchestrator	Evidence self-commit per the close standard	commit-change (evidence scope)	Close

quality-auditor is the single domain agent (1 of the ≤5 budget); the other two are utility. The judge skill carries no Bash — it cannot run mechanics (C4).

Pre-flight

Check	Constraint	Action on Failure
Resolve config + tokens (.garura/core/config.yaml)	—	Hard halt
On the epic's issue branch (issue resolved from branch)	C10-position	Hard halt
Epic eligible: in_delivery + issue_ref; implement's done verdict holds (pieces done, gates pass, steelman PASS)	C1	Hard halt (REC1)
Loop cap not reached (rejections < 3)	C10	Hard halt → escalation record (REC10)

python3 scripts/preflight.py --play validate --config .garura/core/config.yaml \
        --branch "$(git branch --show-current)" --porcelain-file <captured>
python3 scripts/check_ready_validate.py --product-base <product_base> --epic <epic_id> \
        --plan {stm_base}{issue}/plan.yaml \
        --gates {stm_base}{issue}/gates-results.yaml \
        --verdict {stm_base}{issue}/verdict.yaml \
        --status-dir {stm_base}{issue}/validate/status --max-rounds 3

preflight.py returns config facts (stm_base, product_base, evidence_record, the issue from the branch). The epic resolves by id to its entry in the spine epics index ({product_base}/product-os/_spine.yaml); its grounding lives in the epic.md the entry points to. check_ready_validate.py is the eligibility gate AND round resolver: it emits round (prior rejections + 1) and prior_report (the last fix report, the round-N scope source). At the cap it sets escalate: true — present the round history to the human and stop; never start the round (C10/F10). Also capture the two live reads to files now: git status --porcelain > {working}/porcelain-before.txt (SE-12 compares after the run).

{working} is the issue's STM context dir ({stm_base}{issue}/context/) — the baked stop-condition manifest reads the run's artifacts there, relative to the STM root.

Right after the resolver, record the session identity stamp's start marker (#463 — soft-fail, never a halt):

python3 scripts/session_stamp.py --phase start \
    --marker "{stm_base}{issue}/status/session-stamp-validate.json" \
    --cwd "$(pwd)" --branch "$(git branch --show-current)"

Resume check: if {stm_base}{issue}/validate/status/round-<n>-progress.json exists, resume — skip completed steps, reset any in-progress step to pending.

Task DAG

Create ALL tasks immediately after pre-flight — before any domain work. The play owns this DAG; agents must not edit its top-level tasks.

[T1] Resolve blast scope                 blockedBy: []
[T2] Plan checks (KB-grounded)           blockedBy: [T1]
[T3] Grounding check (KB)                blockedBy: [T2]
[T4] Checkpoint (skippable)              blockedBy: [T3]
[T5] Run checks (runner family)          blockedBy: [T4]
[T6] Gates + benchmarks                  blockedBy: [T5]
[T7] Judge findings                      blockedBy: [T6]
[T8] Compute verdict                     blockedBy: [T7]
[T9] Stamp epic                          blockedBy: [T8]
[T10] Render + post report               blockedBy: [T9]
[T11] Verify the run                     blockedBy: [T10]
[T12] Scenario Validation                blockedBy: [T11]
[T13] Close                              blockedBy: [T12]

Mark each task in-progress before its step and completed right after its eval passes. No runtime reordering. On resume, skip completed and reset in-progress to pending.

Workflow

Phase: Scope

Step 1 — Resolve blast scope · Owner: play · Depends on: pre-flight

python3 scripts/resolve_blast_scope.py \
        --reports-dir {stm_base}{issue}/piece-reports \
        --round <round> [--prior-report <prior_report>] \
        --out {working}/scope.json

Round 1 scopes from implement's piece reports (files_modified — the build's recorded claims); round N+1 from the FIX's piece reports plus what the prior report's findings named. The script can only read claims or fail loud — it cannot invent scope. SE-2 (F8/C6): resolve_blast_scope.py exits 0; scope.json.source is implement-claims (round 1) or fix-claims+prior-report (later rounds); its files are non-empty and every entry traces to a piece report or the prior report.

Phase: Plan

Step 2 — Plan checks (KB-grounded) · Owner: quality-auditor · Depends on: Step 1 The agent gathers the epic, the slice's quality + architecture lenses, the profile, the scope, and the repo, and invokes plan-validation-checks:

{
  "task":    "author the check manifest for this epic: detect the stack(s) (arch lens + repo manifests), resolve each to KB-grounded tooling (build, tests, scoped regression per scope.json, lint/static analysis, dependency audit, the scanners the bars demand), map every quality-lens gate (quality.md "## Gates" table) and measure-lens metric (measure.md "## Metrics" table) to a concrete check; code-level security only; round > 1 plans only what the prior report and narrowed scope require; uncovered tool choices become KB-learning-gap proposals, never silent picks. ALSO map the epic's declared surface.type (surface-contract.md) to its ONE required runnable check and add it to the manifest tagged `surface_check: <surface.type>` — a real browser check that opens each must_open artifact (web_dashboard), an HTTP/API call to the declared endpoint (server_api), a run of human_run_target (cli), or the library/service's own tests (library/service_read_model). The plan is INCOMPLETE for a required (user-facing) surface unless it carries that surface_check entry — never omit it for lack of a browser/API probe; the absence of a probe fails the surface, it does not waive it.",
  "inputs":  { "epic_file": "<epic_file>",
               "quality_lens": "<lens_dir>/quality.md",
               "arch_lens": "<lens_dir>/architecture.md",
               "measure_lens": "<lens_dir>/measure.md",
               "profile_path": "<product_base>/product-os/_spine.yaml",
               "scope_file": "{working}/scope.json",
               "repo_root": "<repo_root>",
               "round": "<round>",
               "kb_search": "<skills>/kb-search/scripts/kb_search.py",
               "kb_root": "<knowledge_dir>" },
  "outputs": { "out_dir": "{working}/plan/",
               "manifest": "{working}/plan/checks.yaml",
               "choices": "{working}/plan/check-choices.yaml",
               "proposals": "{working}/plan/proposals/" }
}

SE-4 (F9/C5): every checks.yaml entry's runner is one of the bundled runner classes (java, dotnet, node, frontend, sql, lint, sonar) and no command targets a remote system — the plan is code-level/local by construction; a probe-class check cannot execute because no runner exists for it. SE-14 (F13/C13): the plan is complete against the epic's surface — checks.yaml carries exactly one entry tagged surface_check: <epic.surface.type> for the declared type (for a user-facing type web_dashboard/server_api/cli that entry is the required runnable check: a browser open of each must_open, an HTTP call to the endpoint, or a run of human_run_target; for library/service_read_model it is the own-tests entry). A required surface with no surface_check entry is an INCOMPLETE plan — the check is not waived for lack of a browser/API probe; loop back to Step 2 to add it before running.

Step 3 — Grounding check (KB) · Owner: play · Depends on: Step 2

python3 scripts/check_kb_grounding.py --manifest {working}/plan/check-choices.yaml \
        --proposals-dir {working}/plan/proposals --kb-root <knowledge_dir>

SE-3 (F12/C12): exits 0 — every tool pick and scoping decision carries a KB learning id that resolves or a proposal file that exists; none is invented. On GAP, apply REC12.

Phase: Checkpoint (skippable)

Step 4 — Review the plan · Owner: play · Depends on: Step 3 · Checkpoint (class: standard) Resolve the checkpoint's gate FIRST, per standards/rules/gate-config.md (first match wins: gates.plays.validate → gates.classes.standard → gates.default; absent ⇒ on). Under #467 Batch D gates.plays.validate is off — the human plan-review is redundant because the manifest is fully DERIVED (Step 2 detects the stacks, Step 3 grounds every tool choice in the KB, and any ungroundable pick already halted as a KB-learning-gap at Step 3, C12/F12). The gate governs ONLY the human review; that grounding wall is the machine's and is never gated (C15).

• off — do NOT wait. The precondition that replaces the human review is the machine wall: check-choices.yaml grounded clean at Step 3 — every check traces to a stack-detection result plus a KB-grounded tool choice, no ungrounded pick. Record gate skipped by config (gates.plays.validate) as a Checkpoint Decisions row in the evidence and proceed. If any tool choice is ungrounded and was NOT recorded as a KB-learning-gap, do NOT proceed — apply REC15: halt, run the KB search, record the gap, re-derive, and proceed only when the manifest is fully grounded (this halt is the machine wall, never gated).
• on — present the check plan in plain language: the stacks detected, the checks per bar, the regression scope and its source, any KB gaps or lens-vs-repo mismatches from notes. Skip rule: when check-choices.yaml grounded clean with zero proposals and the plan's notes are empty, skip — post the summary and continue. Otherwise wait for a typed response.

SE-16 (F15/C15): every entry in checks.yaml traces to a stack-detection result plus a KB basis (its check-choices.yaml learning id resolves) OR to a recorded KB-learning-gap proposal under {working}/plan/proposals/ — the off-path precondition holds; a proceed on an ungrounded, unrecorded choice is F15 (REC15). The Step 4 gate resolution (fired, skipped, or its config-skip row) is recorded in the evidence either way — never silent.

Phase: Execute

Step 5 — Run checks (runner family) · Owner: play · Depends on: Step 4

python3 scripts/run_checks.py --manifest {working}/plan/checks.yaml \
        --results-dir {working}/results [--only <prior-findings' check ids, round > 1>]

The orchestrator dispatches each check to its runner as a concurrent read-only fan-out (standards/rules/concurrent-fanout.md): the checks run in a bounded worker pool and JOIN before the summary is built. The three safety conditions hold — each check only READS the code (SE-12 proves the tracked-file set is byte-unchanged), each writes only its own <check_id>.json, and no check consumes another's output; results are collected in manifest order, so summary.json is identical to a serial run. Runners ensure their tool (install when the manifest says how), run it, and emit normalized records. A tool that cannot run is status: error — recorded, never skipped. SE-5 (F5/C4, C7): every result in {working}/results/ was emitted by a runner (each carries raw_log_path; summary.json counts them); no check result was authored by an agent or inline prose; everything that ran, ran to TEST — the manifest carries no acceptance-deploy entry (deploying for human acceptance is /launch's, C7). SE-12 (F4/C2): re-capture git status --porcelain > {working}/porcelain-after.txt; the tracked-file set matches porcelain-before.txt — validate edited no product code (build outputs are untracked; any tracked diff is REC4).

Step 6 — Gates + benchmarks · Owner: play · Depends on: Step 5

python3 scripts/check_gates.py --quality-lens <lens_dir>/quality.md \
        --measure-lens <lens_dir>/measure.md \
        --results-dir {working}/results --out {working}/gates-map.json \
        --product-base <product_base> --epic <epic_id>

SE-11 (F3-partial/C11): every quality-lens gate (read from quality.md's "## Gates" table, by dimension) maps to a captured result and every measure-lens metric (read from measure.md's "## Metrics" table) is captured and, where its Target cell yields a comparator + number, at/above its floor (or at/below its ceiling) — or it is a FINDING in gates-map.json with citation + location (gate-unmeasured, benchmark-below-floor). The floor is the bar; exit 1 with findings is a valid (failing) outcome, not an error. SE-14 (F13/C13) — verdict half: check_gates.py (with --product-base --epic) requires a captured result tagged surface_check: <epic.surface_type> with status pass; a required surface the run did not measure, or a captured surface result whose type does not match the declared surface.type, is a finding (surface-unmeasured / surface-mismatch / surface-failed) in gates-map.json with citation + location — so the verdict step (Step 8) computes fix_required, never validated, on an unmeasured required surface. A browser/API/CLI promise cannot pass on code-level checks alone.

Phase: Judge

Step 7 — Judge findings · Owner: quality-auditor · Depends on: Step 6 The agent reads ONLY the captured artifacts and invokes judge-validation-results:

{
  "task":    "compose the findings from the captured results: extract the exact failing assertion/rule/file:line from each failed check's raw log, deduplicate root causes across tools, flag gamed-looking patterns (collapsed test counts, fresh suppressions) — every finding cites captured output and a location; no speculation, nothing run, nothing fixed",
  "inputs":  { "results_dir": "{working}/results",
               "gates_map": "{working}/gates-map.json",
               "scope_file": "{working}/scope.json" },
  "outputs": { "out_findings": "{working}/findings.yaml" }
}

SE-6 (F2/C3): every entry in findings.yaml carries a citation quoting captured output and a location (file:line or the raw log path) — enforced again mechanically at Step 8, which refuses uncited findings.

Phase: Verdict

Step 8 — Compute verdict · Owner: play · Depends on: Step 7

python3 scripts/compute_verdict.py --summary {working}/results/summary.json \
        --gates-map {working}/gates-map.json --findings {working}/findings.yaml \
        --round <round> --epic-id <epic_id> --out {working}/verdict.json

SE-7 (F3/C8): the verdict is computed mechanically — validated only when every check passed AND zero gate/benchmark findings AND zero judge findings; the script refuses uncited findings (ok=false) rather than weakening the verdict. Record the round: copy verdict.json → {stm_base}{issue}/validate/status/verdict-round-<round>.json (with report set to the report path once rendered) — the next round's gate reads these.

Step 9 — Stamp epic · Owner: play · Depends on: Step 8 Snapshot first, then the one durable model write:

cp <product_base>/product-os/_spine.yaml {working}/spine-before.yaml
python3 scripts/stamp_epic.py --product-base <product_base> --epic <epic_id> \
        --verdict-file {working}/verdict.json --record {working}/verdict.json

On a successful stamp, --record marks the verdict artifact stamped: true in place — the value the stop-condition gate reads at close (#464, Done means D4; a verdict stamped either way counts as done, an unstamped run does not — C14). Then re-copy verdict.json → {stm_base}{issue}/validate/status/verdict-round-<round>.json so the durable round record carries the stamp.

On a validated verdict the stamp also sets surface_verified: true — the surface-parity gate (C13) passed, so the epic records that its required surface was measured and matched (surface-contract.md); /next reads this to distinguish a surface that shipped from surface debt. On fix_required, surface_verified is left untouched.

SE-8 (F6/C8): stamp_epic.py exits 0 and the epic's status now equals the verdict — fix_required actually blocks /launch and re-admits /implement; validated actually unblocks it. SE-9 (F11/C8): Step 11 proves against the snapshot that only status, metadata.version, and (on a pass) surface_verified: true changed.

Phase: Report

Step 10 — Render + post report · Owner: play, then project-orchestrator · Depends on: Step 9

python3 scripts/render_fix_report.py --verdict {working}/verdict.json --round <round> \
        --out-yaml {working}/report.yaml --out-md {working}/report.md

Then dispatch project-orchestrator to post report.md to the epic's tracked issue via manage-issue (comment) — the report is implement's work list and must survive a lost session, like implement's published plan. Update the round record's report field with the report.yaml path. SE-10 (F7/C9): render_fix_report.py exits 0 (it refuses uncited entries); on fix_required every rendered entry names the check id, the citation, and the location; the issue carries the posted comment.

Phase: Verify

Step 11 — Verify the run · Owner: play · Depends on: Step 10

python3 scripts/check_validate.py --verdict {working}/verdict.json \
        --summary {working}/results/summary.json --gates-map {working}/gates-map.json \
        --spine-before {working}/spine-before.yaml --product-base <product_base> --epic <epic_id> \
        --report-yaml {working}/report.yaml --report-md {working}/report.md

SE-1 (F1/C1): holistic re-assert — the run only got here through the eligibility gate (pre-flight exit 0 recorded in the round progress record). SE-13 (F10/C10): the round number in verdict.json is ≤ 3 and, when this round was a third rejection, the close presents the escalation record (all three rounds' findings + fixes) instead of inviting a fourth. check_validate.py exits 0: the verdict matches the recomputed evidence (F3), the stamp matches the verdict (F6), the epic write was surgical (F11), and a failed run carries a complete, cited report (F7).

Phase: Scenario Validation

Step 12 — Scenario evals · Owner: play · Depends on: Step 11

• SCE-1 (S1 — delivery lead): on a clean run, the epic file's status is validated, verdict.json cites every captured check (counts match summary.json), and its findings list is empty.
• SCE-2 (S2 — implementer): on a failing run, the epic status is fix_required, every report.yaml finding carries check id + location, and on the re-run round the stamp flips to validated with total rounds ≤ 3.
• SCE-3 (S3 — security auditor): scan results exist under {working}/results/ with rule ids in their findings' citations; no check ran outside the runner family; the grounding check exits 0 with zero ungrounded tool choices.
• SCE-4 (S4 — product owner): the regression checks' commands trace to scope.json, whose source is recorded claims; each regression failure in the findings carries its check id.
• SCE-5 (S5 — human): at three rejections the next invocation's eligibility gate exits with escalate: true, no round-4 artifacts exist, and the escalation record lists each round's findings and fixes.

Phase: Evidence & Close

Step 13 — Close · Owner: play · Depends on: Step 12 Run the Standard Play Close. /validate is project-scoped (the epic's issue). The close is GATED (the stop-condition step below, #464): the verdict decides COMPLETED vs HALTED. SE-15 (F14/C14): the stop-condition verdict read held before COMPLETED was stamped — D1–D4 all held; a run whose checks ran but whose verdict never stamped closes HALTED (stop_condition_unmet), never COMPLETED. A stamped fix_required counts as done — the gate binds the RUN's completeness, not the epic's quality.

# --- Standard Play Close (canonical; see standards/rules/play-close.md) ---
# Path tokens resolved at pre-flight (resolve here if not already):
#   ltm_project_target  = yq '.ltm.project-target' .garura/core/config.yaml
#   evidence_base, slug:
#     project-scoped play : evidence_base="${stm_base}${issue}/evidence/validate/"  ; slug="#${issue}"
#     product-scoped play : evidence_base="${product_base}_evidence/validate/"      ; slug="${product_slug}"
evidence_template=$(cat "${ltm_project_target}standards/templates/evidence-file.md")
delivery_template=$(cat "${ltm_project_target}standards/templates/delivery-report.md")
ts=$(date -u +%Y%m%d-%H%M%S)
evidence_dest="${evidence_base}${ts}.md"
mkdir -p "$(dirname "$evidence_dest")"
# Session identity stamp (#463) — close phase; start phase ran at pre-flight
session_stamp=$(python3 scripts/session_stamp.py --phase close \
    --marker "${stm_base}${issue}/status/session-stamp-validate.json")
# Stop-condition gate (#464) — Step C0: this play carries a baked manifest, so the
# gate is LIVE. Single-pass play: the Done means is proven here, once.
python3 scripts/check_stop_condition.py \
    --manifest "<play-dir>/stop-condition.yaml" \
    --base "${stm_base}${issue}/" \
    --out "${stm_base}${issue}/status/stop-condition-validate.yaml"
sc_exit=$?   # 0 held · 1 unmet · 2 error

Step C0 — bind the verdict. sc_exit == 0 (held) permits status: COMPLETED. Anything else closes HALTED with exit_reason: stop_condition_unmet and the evidence's Stop Condition section names every unmet clause. An unevaluable verdict is never a pass. The bind is about the RUN, not the epic: a fix_required verdict, stamped, still closes COMPLETED — an unstamped run does not (C14).

Step C1 — Write evidence file. Gated by the resolved evidence.record flag (global + per-play evidence.plays.validate; first match wins, absent ⇒ record). When false, skip the write and record evidence skipped (record=false) in the report's pointer line. Otherwise fill the evidence-file.md slots (play validate, run_id validate-${ts}, the issue, started_at/completed_at, status per C0, exit_reason; artifacts produced: scope.json, checks.yaml + check-choices.yaml, results/ + summary.json, gates-map.json, findings.yaml, verdict.json, the round record, report.yaml/.md, the stamp diff, the stop-condition verdict; step and scenario eval results SE-1…SE-15 / SCE-1…SCE-5; the Step 4 checkpoint decision or skip (or its config-skip row); the session identity stamp fields from $session_stamp (#463): session_id, ledger_file, ledger_start_offset, ledger_end_offset (null when unresolved — never blocks the close); and stop_condition per C0 with the Stop Condition section filled) and write to $evidence_dest. Do NOT hand-author the body. Then dispatch repo-orchestrator for the evidence self-commit when evidence was recorded (the WORK is never committed by this play — the branch is /launch's to close).

Step C2 — Render delivery report. Also render the Next line: resolve this play in standards/rules/pipeline-next.md and emit **Next:** /<command> — <why>. Or run /next to see all recommended actions. (only /next pointer, or omit, when the mapped command is null), per play-close.md. Fill the delivery-report.md slots and output the report: ## validate Round <round> — #${issue} (the verdict, plainly: validated → "/launch is next"; fix_required → "the fix report is on the issue — run /implement to fix exactly what it names"; escalation → the three-round history), the Run Summary table, the Pipeline Steps table from the task DAG, the Artifacts Produced table, Next Steps, and a pointer to $evidence_dest. Always emitted; never gated.

# --- end Standard Play Close ---

Scenario Validation

Scenario	Persona	Eval
S1 — happy path	delivery lead	SCE-1
S2 — the fix loop	implementer	SCE-2
S3 — security pass	security auditor	SCE-3
S4 — nothing shipped broke	product owner	SCE-4
S5 — escalation	human	SCE-5

Recovery

For	Trigger	Direction	Handoff
F1	implement's done verdict missing or incomplete at pre-flight	halt; the epic must finish /implement first	human
F2	a finding has no mechanical citation	rebuild the finding from captured results; re-run its check if needed; drop what cannot be cited	autonomous
F3	the verdict contradicts captured evidence	recompute the verdict mechanically from the result files — never hand-stamp	autonomous
F4	validate edited product code	revert the edit; record it as a finding routed to implement	autonomous
F5	an agent ran mechanics or judged without captured results	re-dispatch with the script doing the work; the agent re-judges over its output	autonomous
F6	a failed run carries no fix_required stamp	apply the stamp via the surgical writer before close	autonomous
F7	a fix report too vague to act on	regenerate entries with check id + location from captured results	autonomous
F8	blast scope not drawn from recorded change claims	rebuild the scope from the recorded claims; re-run regression on it	autonomous
F9	probing beyond code-level	discard those results; re-run the security pass code-level only	autonomous
F10	the cap reached without escalation	stop the loop; produce the escalation record with round history	human
F11	the epic record touched beyond the stamp	restore the record and re-apply only the surgical stamp	autonomous
F12	an ungrounded scan choice with no gap recorded	run the KB search; if still uncovered, record the gap proposal, then proceed	autonomous
F13	the required surface check for the epic's surface.type was not run, or the captured surface result does not match the declared type	stamp fix_required naming the unmeasured/mismatched surface (the surface.type and the must_open/human_run_target it owed), and re-admit /implement to restore it so the next round can measure it	autonomous
F14	the close would stamp COMPLETED without the stop-condition verdict held	evaluate the stop condition, surface the unmet clauses, and close HALTED (stop_condition_unmet) until they are fixed — an unevaluable verdict is never a pass	autonomous
F15	with the plan-review gate off, the manifest reached the check-execution step carrying an ungrounded tool choice that was neither KB-matched nor recorded as a KB-learning-gap	halt before the checks run; run the KB search for the uncovered choice, record the KB-learning-gap proposal, re-derive the manifest, and proceed only when every check is fully grounded	autonomous

Pause and Resume

Steps run top to bottom. On entry, resolve config, capture the live reads, run the eligibility + round gate, check the round progress marker, skip completed steps, reset any in-progress step to pending, and continue. The epic snapshot is captured at Step 9 (the gated stamp step) and preserved on resume, so the surgical comparison always diffs against true pre-stamp state. A fresh start with no marker runs everything and creates the marker at Step 1. Each round is its own progress marker; a resumed round never re-counts itself.

Compilation Metadata

Field	Value
fingerprint	sha256:01bbfe25700849b61f01cc3d0ee6390e2578db62c704560ea22831e01cea3ac8 (of reference/ice.md)
compiled_by	play-editor (#466 Batch B; #467 Batch D — Step 4 gate off + C15/F15/REC15/SE-16)
pipeline_position	none
position_exception	middle of the execute pipeline (implement → validate → launch) — runs on the epic branch /implement started; the close chain belongs to /launch after human sign-off (#434, decisions 20 + 24)
structural_constraints	C2-part (finder/fixer split — no code-edit step exists), C4 (judge skill carries no Bash; mechanics only via runners), C7 (no acceptance-deploy step — run-to-test only), C9 (remediation only via the posted report; no builder dispatch), C10-position (execute middle — no head, no close)
workflow_structure	A (Step 4 checkpoint gate off under #467 Batch D — the human plan-review is replaced by the KB-grounding wall, C15; still skippable when on and the plan grounds clean); single-pass — no loop, no cap; gated close
stop_condition	stop-condition.yaml (D1–D4), gate live at Step C0
domain_agents	1 (quality-auditor)
utility_agents	2 (project-orchestrator, repo-orchestrator)
skills_used	plan-validation-checks, judge-validation-results, kb-search, manage-issue
standards_consumed	standards/rules/gate-config.md (Step 4 switch); standards/rules/play-close.md; standards/rules/pipeline-next.md
scripts	13 + 8 runners (preflight.py, check_ready_validate.py, resolve_blast_scope.py, run_checks.py, check_gates.py, check_kb_grounding.py, compute_verdict.py, stamp_epic.py, render_fix_report.py, check_validate.py, session_stamp.py — session identity stamp, check_stop_condition.py — Done-means gate, runners/runner_common.py + 7 runner*.py)
step_evals	16 (SE-1…SE-16; every failure condition covered; SE-14 covers the surface contract F13/C13 at the plan and verdict halves; SE-15 covers the gated close F14/C14; SE-16 covers the gate-off manifest-grounding precondition F15/C15)
scenario_evals	5 (SCE-1…SCE-5)
recovery_entries	15 (one per failure condition; 13 autonomous / 2 human)

Direct-edit deviation note (#434, spine + grounding model)

Non-intent edit: the epic moved from a per-epic epic.yaml file to an entry in the spine epics index (product-os/_spine.yaml) plus an epic.md grounding doc. The mechanism scripts were retargeted: check_ready_validate.py reads the epic entry by id from the spine; check_gates.py reads the epic's surface_type from the spine entry (was epic.yaml surface.type); stamp_epic.py makes the in_delivery → validated/fix_required flip (plus surface_verified on a pass) as a surgical write to the spine epics entry (--product-base --epic); the surgical-write proof now diffs the epic entry between a pre-stamp spine snapshot (--spine-before) and the live spine (check_validate.py). Prose invocations updated to match. No constraint/failure/scenario/eval text changed — the verdict logic, the surface gate, and the surgical-write guarantee are identical; only the storage moved. reference/ice.md and the fingerprint are unchanged.

Direct-edit deviation note (#434, validate-readers-spine-migration): check_gates.py's gate + benchmark DEFINITIONS now read the lens GROUNDING DOCS, not YAML: quality gates parse quality.md's "## Gates" table (gate id = dimension slug; captured check_id matches the slug), and benchmarks parse measure.md's "## Metrics" table (each metric is presence-required, and where its Target cell yields a comparator + number it is also compared as a floor/ceiling). The old standalone profile.yaml benchmark source is retired — the structured floor/ceiling numbers it carried now live as prose targets in the measure lens, so a tolerant comparator parse stands in (prose-only targets are presence-checked). CLI: --quality-lens now takes quality.md, --profile is replaced by --measure-lens <measure.md>; the plan-validation-checks contract points at the .md lenses + the spine. The gate LOGIC (every declared bar maps to a captured result; an unmeasured bar is not a passed bar) is identical — only the definition source moved from YAML to the linted, eval-gated grounding docs. reference/ice.md and the fingerprint are unchanged.

Direct-edit deviation note (#468 Stage 5, concurrent fan-out): run_checks.py now runs its per-tool check runners as a concurrent read-only fan-out — a bounded worker pool that joins before the summary is built (standards/rules/concurrent-fanout.md) — instead of a serial loop. The three safety conditions hold (each check reads only the code, writes only its own <check_id>.json, and depends on no other), and results are collected in manifest order, so summary.json is byte-identical to a serial run; a --max-parallel 1 flag forces the old serial behavior. Execution-timing change only: no constraint/failure/scenario/eval touched, the finder-not-fixer guarantee (SE-12) is unchanged, and reference/ice.md and the fingerprint stand. A duplicate-check_id guard was added because concurrent runners must not race on the same result file.