ouroboros-ai 0.40.1.dev2__tar.gz → 0.40.2.dev2__tar.gz

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
     {
       "name": "ouroboros",
       "description": "Transform vague ideas into validated seed specifications through Socratic questioning and ontological analysis",
-      "version": "0.40.1",
+      "version": "0.40.2",
       "author": {
         "name": "Q00",
         "email": "jqyu.lee@gmail.com"

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ouroboros",
-  "version": "0.40.1",
+  "version": "0.40.2",
   "description": "Self-improving AI workflow system. Crystallize requirements before execution with Socratic interview, ambiguity scoring, and 3-stage evaluation.",
   "author": {
     "name": "Q00",

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ouroboros-ai
-Version: 0.40.1.dev2
+Version: 0.40.2.dev2
 Summary: Specification-first workflow engine for AI coding agents. Works with Claude Code and Codex CLI.
 Author-email: Q00 <jqyu.lee@gmail.com>
 License-File: LICENSE

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/docs/README.md

@@ -30,6 +30,7 @@ replayable execution contract on your choice of runtime backend.
 - [Configuration Reference](./config-reference.md) - All `config.yaml` options and environment variables
 - [Agent OS Profile Taxonomy](./agentos/profile-taxonomy.md) - Locked 4-slot vocabulary (`runtime_backend`, `stage_runtime`, `llm_profile`, `provider_profile`) for the historically-overloaded "profile" concepts; tiebreaker for #573
 - [AgentOS Release Readiness](./agentos/release-readiness.md) - #961-aligned release gate, verification pack, and open issue triage
+- [AgentOS Issue Sequencing Graph](./agentos/issue-sequencing-graph.md) - Non-normative, dated #961/#1256 snapshot of track sequencing, merged-PR evidence, and recommended work order
 
 ### API Reference
 
@@ -55,6 +56,7 @@ replayable execution contract on your choice of runtime backend.
 - [Agent OS Kernel Terminology](./contributing/agent-os-kernel-terminology.md) - Locked vocabulary for `AgentRuntimeContext`, `ControlPlane`, `ControlContract`, `Directive`, `ControlBus`, and `IOJournal`
 - [ControlContract](./contributing/control-contract.md) - Control-plane schema, terminality, replay, and idempotency invariants
 - [Testing Guide](./contributing/testing-guide.md) - Writing and running tests
+- [Verifier Evidence Policy](./contributing/verifier-evidence-policy.md) - Core verifier boundary: avoid runner-specific parsers, classify evidence-form mismatches, and preserve anti-fabrication semantics
 - [Key Patterns](./contributing/key-patterns.md) - Result type, immutability, event sourcing, protocols
 - [Findings Registry](./contributing/findings-registry.md) - Documentation audit findings registry
 - [Issue Quality Policy](./contributing/issue-quality-policy.md) - Quality bar for actionable issues and PRD-lite feature requests

ouroboros_ai-0.40.2.dev2/docs/agentos/issue-sequencing-graph.md

@@ -0,0 +1,241 @@
+# AgentOS Issue Sequencing Graph
+
+> Snapshot documentation imported from #1291. This page is informational only; #961 remains the canonical AgentOS roadmap SSOT, and #1256 remains the product-policy SSOT for the `ooo auto` closure ladder.
+
+Snapshot: 2026-05-29 18:40 KST.
+
+This document is a point-in-time triage view for the AgentOS issue set. It is
+not a new source of truth. GitHub issues, especially #961 and #1256, remain the
+normative records. Use this page to see which recently merged PRs resolved
+which roadmap tracks, which open issues are still intentionally open, and what
+order the remaining work should follow.
+
+## Source Of Truth
+
+- #961 is the AgentOS roadmap sequencing SSOT.
+- Track B's implementation merge train is reconciled. #1256 owns the remaining
+  product-policy decisions, but they are not open autonomous implementation
+  work.
+- #939 remains the plugin lifecycle umbrella after its docs/schema slices.
+- The GitHub open issue list is the closure authority. At this snapshot there
+  are 20 open issues, and none has enough merged evidence to close safely.
+
+## Sequencing Graph
+
+```mermaid
+flowchart TD
+    LegendMeta["Meta / RFC SSOT"]
+    LegendOpen["Open: design or implementation remains"]
+    LegendMerged["Merged PR / delivered slice"]
+    LegendClosed["Closed issue"]
+    LegendBacklog["Backlog / later design"]
+
+    I961["#961 AgentOS roadmap SSOT"]
+
+    I961 --> A["Track A: fat harness COMPLETE"]
+    A --> A920["#920 closed"]
+    A --> A925["#925 closed"]
+    A --> A960["#960 closed"]
+    A --> A1282["#1282 merged: health readiness snapshot"]
+
+    I961 --> B["Track B: ooo auto IMPLEMENTATION RECONCILED"]
+    B --> I1157["#1157 auto vision SSOT"]
+    I1157 --> I1256["#1256 Product-or-Die RFC"]
+    I1256 --> B1257["#1257 closed by #1269-#1272"]
+    I1256 --> B1260["#1260 merged: first EventStore slice"]
+    I1256 --> I1254["#1254 open: all auto state transitions"]
+    I1256 --> I1258["#1258 open: perf regression approval/design"]
+    I1256 --> I1170["#1170 open: canonical product matrix"]
+    I1256 --> I1263["#1263 open: aggressive autofill design"]
+    I1256 --> I1234["#1234 open: verifier pipeline policy"]
+    B --> B1281["#1281 merged: checkpoint isolation"]
+    B --> B1284["#1284 merged: interview progress relay"]
+    B --> B1286["#1286 merged: detached job UX"]
+    B --> B1289["#1289 merged: #1283 residue"]
+    B --> B1290["#1290 merged: #1285 residue"]
+
+    I961 --> C["Track C: AgentOS substrate / plugins"]
+    C --> C946["#946 closed"]
+    C --> C956["#956 closed"]
+    C --> C1276["#1276 merged: artifact/state hook contract"]
+    C --> C1277["#1277 merged: plugin v0.4 schema"]
+    C --> I939["#939 open: remaining lifecycle hooks/audit work"]
+
+    I961 --> D["Design backlog"]
+    D --> I518["#518 AgentProcess lifecycle"]
+    D --> I614["#614 external guidance contract"]
+    D --> I615["#615 reasoning budget policy"]
+    D --> I813["#813 Symposium primitive"]
+    I813 --> I814["#814 verdict envelope"]
+    I813 --> I815["#815 stateful debate policy"]
+    I813 --> I816["#816 evaluate consensus migration"]
+    I813 --> I819["#819 persistent transcript"]
+    D --> I831["#831 interview MCP hang watchlist"]
+    D --> I1139["#1139 Pi runtime bridge"]
+    D --> I1239["#1239 reference-aware interviews"]
+
+    classDef meta fill:#fef3c7,stroke:#d97706,color:#111827,stroke-width:2px
+    classDef open fill:#fee2e2,stroke:#dc2626,color:#111827,stroke-width:2px
+    classDef merged fill:#dcfce7,stroke:#16a34a,color:#111827,stroke-width:2px
+    classDef closed fill:#e0f2fe,stroke:#0284c7,color:#111827,stroke-width:2px
+    classDef backlog fill:#f3e8ff,stroke:#7e22ce,color:#111827,stroke-width:2px
+    classDef track fill:#f8fafc,stroke:#64748b,color:#111827,stroke-width:1px
+
+    class LegendMeta,I961,I1157,I1256 meta
+    class LegendOpen,I1254,I1258,I1170,I1263,I1234,I939 open
+    class LegendMerged,A1282,B1260,B1281,B1284,B1286,B1289,B1290,C1276,C1277 merged
+    class LegendClosed,A920,A925,A960,B1257,C946,C956 closed
+    class LegendBacklog,I518,I614,I615,I813,I814,I815,I816,I819,I831,I1139,I1239 backlog
+    class A,B,C,D track
+```
+
+## Closure Verdict
+
+No currently open issue is safe to close from the available evidence.
+
+The merged PRs have resolved important slices, but the open issues are still in
+one of these states:
+
+- Meta or RFC SSOT that must stay open while downstream tracks converge.
+- Partially implemented umbrella with remaining scoped work.
+- Needs-design or needs-approval item that requires a human decision before
+  implementation can be declared complete.
+- Backlog design issue unrelated to the recent auto/detached-job merge train.
+
+Already closed items with clear closure evidence include #920, #925, #946,
+#956, #960, #1257, #1267, #573, #575, and #476.
+
+## `ooo auto` Release Position
+
+Track B is release-ready as an implementation track, but #1256 should stay open
+as a policy/RFC umbrella.
+
+What this release can safely claim:
+
+- `ooo auto --complete-product` has a real Interview -> Seed -> Run -> Ralph
+  product chain, with foreground completion when Ralph reaches a terminal
+  success and tracked detached background work when ownership moves out of the
+  client wait budget.
+- Detached auto is not treated as failure: CLI/MCP status, wait, and result
+  surfaces expose stable handles and polling guidance.
+- The R3 failure class from #1256/#1257 is fixed: an interview phase deadline no
+  longer terminally reports `interview_phase_deadline` BLOCKED. It routes
+  through `partial_seed_from_evidence`, emits `runtime.deadline.interview.fired`
+  before `auto.product.partial_emitted`, and surfaces
+  `partial_product=True` with unresolved slots as next-step hints.
+- Degraded partial Seeds relax ambiguity/open-gap blockers only for recoverable
+  missing evidence. Safety blockers and explicit human-required ledger BLOCKED
+  entries still block.
+- Canonical `cli-todo` product-complete shape exists and validates fixture
+  shape in CI; the live canonical run remains manual opt-in because it consumes
+  real LLM/runtime budget.
+
+What this release should not over-claim:
+
+- #1256 Product-or-Die is not fully adopted as a global invariant. The codebase
+  still has legitimate BLOCKED states for safety, missing structural goal,
+  failed run handoff, Ralph terminal failure, and other recoverable operator
+  states.
+- #1263 aggressive LLM auto-fill is not implemented as a general policy. Safe
+  defaults and degraded partial Seeds exist, but "auto-fill every gap with LLM
+  inference" remains owner-approved future policy.
+- #1254 unified EventStore coverage is only partially implemented. The
+  interview/deadline/partial-product path has typed evidence, but the RFC's
+  "every transition in every phase" contract is still umbrella work.
+- #1258 throughput/per-round regression remains approval-gated. Guardrails and
+  R-run lint exist, but the release should not claim improved auto throughput.
+- #1170 I6 product reality is represented by the canonical harness and
+  `cli-todo` shape/live-run path, not by a full 3-task product matrix.
+
+Focused verification on this snapshot:
+
+```text
+uv run pytest tests/unit/auto/test_pipeline_partial_product_terminal.py \
+  tests/integration/auto/test_interview_deadline_partial_product_regression.py \
+  tests/canonical/test_canonical.py \
+  tests/unit/mcp/tools/test_start_auto.py \
+  tests/unit/cli/test_job_command.py -q
+
+82 passed, 1 skipped
+```
+
+## Recently Merged Evidence
+
+| PR | Merged | Track | Delivered behavior | Disposition |
+| --- | --- | --- | --- | --- |
+| #1290 | 2026-05-29 09:30 UTC | Track B | Carried the linked-stream residue from closed-unmerged #1285 into the branch that actually merged. This preserves the intended stream/attachment behavior without depending on the abandoned PR. | Detached stream residue is merged; no open issue closes from this alone. |
+| #1289 | 2026-05-29 09:31 UTC | Track B | Carried the auto-dispatch residue from closed-unmerged #1283 into the branch that actually merged. This keeps CLI/MCP detached auto dispatch aligned after the superseding UX PR. | Dispatch residue is merged; #1283 itself stays closed-unmerged. |
+| #1286 | 2026-05-29 06:12 UTC | Track B | Improved detached auto job tracking UX so auto runs are treated as tracked background work rather than terminal failures when the client stops waiting. | Main detached async UX slice is merged; follow-up residue landed in #1289/#1290. |
+| #1284 | 2026-05-29 06:12 UTC | Track B | Relays hidden auto interview questions into progress output so users can see what the background interview is asking instead of treating auto as a black box. | Observability UX improved; broader EventStore lifecycle coverage remains #1254. |
+| #1282 | 2026-05-29 07:03 UTC | Track A/C | Added the AgentOS health-readiness table and release-readiness triage snapshot. It records that Track A wiring and baseline gates are complete, then separates remaining Track B/C/D follow-ups from the completed fat-harness lane. | Snapshot merged; #961 remains the live roadmap SSOT. |
+| #1281 | 2026-05-29 06:46 UTC | Track B | Isolates coding sessions with checkpoint commits so `ooo auto` can preserve intermediate work and avoid losing state across detached or resumed coding sessions. | Reliability slice merged; remaining Track B items are design/RFC-gated, not open implementation PR work. |
+| #1280 | 2026-05-28 16:18 UTC | Track B | Cleans up synchronous `complete-product` runs so completed auto sessions do not leave misleading active/blocked residue. | Cleanup merged; detached and background semantics were completed later by #1286/#1289/#1290. |
+| #1279 | 2026-05-28 16:17 UTC | Track B | Requires terminal run evidence before Ralph/evolution handoff, preventing premature evolution loops when auto has not actually produced terminal execution evidence. | Guardrail merged; canonical product matrix remains #1170. |
+| #1278 | 2026-05-28 15:28 UTC | Track B | Closes backend-ready fallback resumes, so fallback-ready sessions can resume through the intended terminal path instead of appearing stranded. | Resume fix merged; broader auto ladder semantics are covered by #1256. |
+| #1277 | 2026-05-28 15:03 UTC | Track C | Adds plugin schema v0.4 plus tool-call hook type promotion, advancing #939's lifecycle hook contract from prose toward typed plugin metadata. | Schema slice merged; #939 remains open for write-side hook/audit execution. |
+| #1276 | 2026-05-28 14:10 UTC | Track C | Documents the artifact/state hook contract for plugins, including what hook producers and consumers can rely on. | Contract docs merged; implementation slices remain under #939. |
+| #1275 | 2026-05-28 14:10 UTC | Track C | Locks the profile taxonomy terminology for #573, removing ambiguity around the overloaded "profile" vocabulary. | #573 is closed. |
+| #1274 | 2026-05-28 14:10 UTC | Track C | Locks ControlJournal option A for #575, settling that design choice for the runtime contract. | #575 is closed. |
+| #1273 | 2026-05-28 14:10 UTC | Track C | Records #476 as implemented in the runtime contract snapshot so the earlier runtime-contract issue has a durable doc anchor. | #476 is closed. |
+| #1272 | 2026-05-28 14:58 UTC | Track B | Adds canonical regression coverage for the #1257 closure ladder, making the ladder behavior testable instead of only described. | Completes the #1257 ladder PR train. |
+| #1271 | 2026-05-28 14:52 UTC | Track B | Makes degraded seed output a partial-product terminal state, so imperfect evidence can still yield a usable artifact instead of BLOCKED. | Part of #1257 closure; #1257 is closed. |
+| #1270 | 2026-05-28 14:36 UTC | Track B | Converts interview deadline expiry into closure-ladder behavior, avoiding terminal BLOCKED when enough partial evidence exists. | Part of #1257 closure; #1257 is closed. |
+| #1269 | 2026-05-28 14:11 UTC | Track B | Adds the `partial_seed_from_evidence` substrate used by the closure ladder to synthesize a partial Seed from collected interview evidence. | Foundation for #1270-#1272; #1257 is closed. |
+| #1268 | 2026-05-28 14:10 UTC | Track C | Migrates Codex setup profiles to profile-v2, aligning setup/runtime naming with the locked taxonomy. | #1267 is closed. |
+| #1260 | 2026-05-28 05:02 UTC | Track B | Wires the first `ooo auto` interview lifecycle events to EventStore, proving the path but only for the initial event family. | First slice merged; #1254 remains open for all auto state transitions. |
+| #1259 | 2026-05-28 05:01 UTC | Track B | Adds CI/perf-budget lint around the R-run section so future auto evidence must include timing discipline. | Guardrail merged; the #1258 wall-clock regression itself remains open. |
+| #1238 | 2026-05-26 16:52 UTC | Track B | Makes the verifier tolerate safe Gradle/Maven evidence cases and quote variants that were falsely flagged by #1234. | Safe subset merged; strict pipeline/tail policy remains open in #1234. |
+
+## Open Issue Disposition
+
+| Issue | Track | Current disposition | Next decision |
+| --- | --- | --- | --- |
+| #961 | Meta | AgentOS roadmap SSOT; intentionally open. | Keep as the single coordination point until AgentOS release gates close. |
+| #1256 | B | Product-or-Die RFC; intentionally open after implementation reconciliation. | Keep as policy SSOT; do not start more autonomous Track B implementation without explicit approval. |
+| #1157 | B | Auto vision SSOT now subordinate to #1256. | Keep open as meta vision until maintainers decide the product-policy lane is done. |
+| #1254 | B | Partially addressed by #1260; expanded by #1256 to all auto state transitions. | Design-gated observability umbrella, not a current release implementation blocker. |
+| #1258 | B | Regression guardrails landed in #1259; the broader perf policy remains needs-design/needs-approval. | Treat as approval-gated policy/perf work, not an active autonomous fix. |
+| #1263 | B | Design item for aggressive auto-fill after non-converging interviews. | Defer until maintainers explicitly accept the policy and scope. |
+| #1170 | B | Canonical harness/product matrix remains a meta slice. | Keep as future acceptance policy unless #1256 promotes a concrete implementation task. |
+| #1234 | B | Safe subset landed in #1238, but pipe/tail policy remains unresolved. | Decide strict matcher policy for unprotected pipelines. |
+| #939 | C | Umbrella remains open after #1276/#1277. | Finish write-side artifact/state substrate and remaining hook/audit slices. |
+| #518 | D | Long-lived AgentProcess lifecycle design item. | Keep for later runtime lifecycle planning. |
+| #614 | D | External guidance contract design item. | Define project/user skill guidance boundaries. |
+| #615 | D | Runtime reasoning budget policy design item. | Define budget/cognitive-effort policy and enforcement surface. |
+| #813 | D | Symposium primitive RFC. | Pick naming/envelope/adoption plan before child migration. |
+| #814 | D | Child of Symposium-style dispatch direction. | Standardize verdict envelope after #813 direction. |
+| #815 | D | Debate-pool policy question. | Decide whether stateful agents may join debates. |
+| #816 | D | Needs-info migration from score average to debate. | Revisit after #813/#814 policy is accepted. |
+| #819 | D | Persistent transcript for multi-persona debate. | Revisit after the debate primitive is accepted. |
+| #831 | D | Watchlist, not an active release blocker. #834/#837/#972 added contract hygiene and diagnostics, but the original trace likely crossed into Claude Code/SDK behavior. | Do not spend more harness-engineering effort unless it reproduces again on the supported MCP interview path; disclose or reopen as active only on a fresh failure. |
+| #1139 | D | Pi support via runtime bridge. | Keep as integration design work, separate from release readiness. |
+| #1239 | D | Reference-aware interview adapters RFC. | Define adapter/glossary provenance contract before implementation. |
+
+## Recommended Work Order
+
+1. Keep #961 and #1256 open as control issues. Do not close them for slice PRs.
+2. Treat the detached async-running lane as merged through #1286, #1289, and
+   #1290. There is no remaining open issue to close for #1283/#1285 because
+   those PRs were closed unmerged and their residue landed in follow-up PRs.
+3. Treat Track B as implementation-reconciled. Do not keep digging there just
+   because #1157/#1170/#1256/#1254/#1258/#1263 remain open; those are
+   meta/RFC/design/approval surfaces. Start new Track B implementation only if
+   a maintainer turns one into a concrete accepted slice.
+4. Finish Track C after the write-side substrate is unblocked:
+   - Remaining #939 hook dispatch, permissions, and audit slices.
+5. Keep Track D as a deliberate design backlog. It should not block the current
+   AgentOS release-readiness lane unless #961 reclassifies one of those issues.
+   For #831 specifically, prefer watchlist/disclosure over more harness work
+   unless a fresh supported-path hang reproduces.
+
+## Practical Triage Rule
+
+When a new PR lands, update this page only if it changes one of these facts:
+
+- An open issue becomes safely closable.
+- A meta issue delegates ownership to a different SSOT.
+- The recommended work order changes.
+- A previously deferred Track B/C item gets a concrete implementation plan.
+
+Otherwise, leave #961 as the active SSOT and avoid creating duplicate roadmap
+state.

ouroboros_ai-0.40.2.dev2/docs/contributing/verifier-evidence-policy.md

@@ -0,0 +1,80 @@
+# Verifier Evidence Policy
+
+This policy governs fixes to the atomic verifier and fat-harness evidence
+boundary.
+
+The core verifier must stay domain- and language-agnostic. It decides whether
+typed evidence is backed by the runtime transcript. It must not grow bespoke
+parsers for each test runner, language ecosystem, framework, or report format.
+
+## Core rule
+
+When valid work appears to be rejected, first classify the failure shape:
+
+| Shape | Meaning | Correct direction |
+| --- | --- | --- |
+| No runtime transcript evidence exists for the claim | The claim may be fabricated. | Keep or use `FABRICATION_SUSPECTED`. |
+| Related runtime work exists, but the evidence form cannot prove the claim | The evidence contract is mismatched. | Use `EVIDENCE_FORM_MISMATCH` with retry guidance. |
+| The evidence needs runner-specific interpretation | The core verifier is the wrong layer. | Add a profile/adapter-level contract or require a runner-agnostic proof form. |
+
+## Do not add runner-specific parsers to core
+
+Avoid fixes that teach `parallel_executor.py` to understand one ecosystem's
+result format, such as JUnit XML, pytest JUnit XML, TAP, Go test JSON, Vitest
+JSON, Maven Surefire XML, or Gradle-specific report layouts.
+
+Those fixes make one stack pass while creating an implicit obligation to support
+every other stack in the same core path. They also make anti-fabrication
+semantics depend on language-specific parsing details that the core verifier
+cannot own consistently.
+
+## Preferred evidence forms
+
+For test commands whose output is filtered or paged, the preferred proof is a
+runner-agnostic command contract:
+
+```sh
+set -o pipefail && <test command> 2>&1 | tail -100
+```
+
+`pipefail` preserves the failing status of the left-hand test command even when
+the right-hand output filter succeeds. Without it, a filtered pipeline is a real
+transcript event but not a clean command proof.
+
+For richer result formats, prefer one of these approaches:
+
+- Emit a typed, runner-neutral proof field from a profile or adapter that owns
+  that runner.
+- Keep the core verdict as `EVIDENCE_FORM_MISMATCH` and retry with clearer
+  evidence.
+- Promote a new cross-runner evidence contract only after it has an explicit
+  design issue and acceptance criteria across multiple ecosystems.
+
+## Failure class semantics
+
+`FABRICATION_SUSPECTED` is reserved for claims with no supporting runtime event
+or artifact reference. It should not be used when the transcript clearly shows
+related work but the evidence shape is contract-incompatible.
+
+`EVIDENCE_FORM_MISMATCH` means:
+
+- related runtime work is visible;
+- the current evidence form cannot prove the typed claim;
+- retrying with a contract-compliant proof form is reasonable;
+- the verifier must still reject the claim until that proof form exists.
+
+This distinction keeps issue triage honest: the implementer did not necessarily
+invent work, but the harness still cannot accept the evidence.
+
+## Review checklist
+
+Before accepting a verifier-evidence fix, ask:
+
+1. Does this add language-, framework-, or runner-specific parsing to core?
+2. Could the same pattern appear in another ecosystem tomorrow?
+3. Is the fix preserving anti-fabrication semantics, or merely making one report
+   format pass?
+4. Would `EVIDENCE_FORM_MISMATCH` plus retry guidance be the correct smaller
+   response?
+5. If structured parsing is required, is it owned by a profile/adapter or a
+   cross-runner evidence contract rather than by the core verifier?

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/skills/setup/SKILL.md

@@ -282,7 +282,7 @@ A backup will be created: CLAUDE.md.bak
 **If "Preview first", show:**
 ````markdown
 <!-- ooo:START -->
-<!-- ooo:VERSION:0.40.1 -->
+<!-- ooo:VERSION:0.40.2 -->
 # Ouroboros — Specification-First AI Development
 
 > Before telling AI what to build, define what should be built.

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/src/ouroboros/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.40.1.dev2'
-__version_tuple__ = version_tuple = (0, 40, 1, 'dev2')
+__version__ = version = '0.40.2.dev2'
+__version_tuple__ = version_tuple = (0, 40, 2, 'dev2')
 
 __commit_id__ = commit_id = None

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/src/ouroboros/orchestrator/failure_taxonomy.py

@@ -28,6 +28,9 @@ class FailureClass(StrEnum):
     Members:
         EVIDENCE_MISSING: Leaf could not emit a parseable / validated
             evidence record (covers both parse errors and H2 rejections).
+        EVIDENCE_FORM_MISMATCH: Leaf executed related work, but the evidence
+            shape cannot prove it under the contract (for example an
+            unprotected output-filter pipeline for a test command).
         FABRICATION_SUSPECTED: Verifier flagged claims about files,
             symbols, or sources that do not exist. Verifier sets this
             via VerifierVerdict.failure_class.
@@ -43,6 +46,7 @@ class FailureClass(StrEnum):
     """
 
     EVIDENCE_MISSING = "EVIDENCE_MISSING"
+    EVIDENCE_FORM_MISMATCH = "EVIDENCE_FORM_MISMATCH"
     FABRICATION_SUSPECTED = "FABRICATION_SUSPECTED"
     SCOPE_CREEP = "SCOPE_CREEP"
     STALL = "STALL"
@@ -74,6 +78,14 @@ _POLICY_TABLE: dict[FailureClass, RecoveryPolicy] = {
             "verifier feedback already names the missing/rejected fields."
         ),
     ),
+    FailureClass.EVIDENCE_FORM_MISMATCH: RecoveryPolicy(
+        action=RecoveryAction.RETRY,
+        rationale=(
+            "Leaf ran related work, but its evidence shape cannot prove the "
+            "claim; retry with contract-compliant evidence such as pipefail "
+            "for output-filtered test commands."
+        ),
+    ),
     FailureClass.FABRICATION_SUSPECTED: RecoveryPolicy(
         action=RecoveryAction.ESCALATE_MODEL,
         rationale=(

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/src/ouroboros/orchestrator/parallel_executor.py

@@ -692,6 +692,86 @@ def _runtime_messages_support_command_claim(
     return any(_runtime_message_supports_command_claim(value, message) for message in messages)
 
 
+def _runtime_messages_have_masked_test_command_form(
+    value: str,
+    messages: tuple[AgentMessage, ...],
+) -> bool:
+    """Return True when a test command claim matches only after unsafe plumbing.
+
+    This deliberately does NOT prove the command claim. It distinguishes a real
+    transcript shape that failed the evidence contract (for example a test run
+    piped through ``tail`` without ``set -o pipefail``) from a fabrication where
+    no related test command appears at all.
+    """
+    claim_invocation = _test_command_invocation(value)
+    if claim_invocation is None:
+        return False
+    for message in messages:
+        if message.tool_name != "Bash":
+            continue
+        for runtime_command in _runtime_message_command_values(message):
+            if not _has_trailing_output_filter_pipeline(runtime_command):
+                continue
+            runtime_invocation = _test_command_invocation_allowing_output_plumbing(runtime_command)
+            if runtime_invocation is None:
+                continue
+            if runtime_invocation == claim_invocation or runtime_invocation.startswith(
+                claim_invocation + " "
+            ):
+                return True
+    return False
+
+
+def _runtime_messages_have_masked_test_command_for_test_claim(
+    *,
+    value: str,
+    messages: tuple[AgentMessage, ...],
+    task_cwd: str | None,
+) -> bool:
+    """Return True when a rejected test claim depends on masked test output.
+
+    This is diagnostic only. It lets the verifier classify a dependent
+    ``tests_passed`` failure with the same evidence-form mismatch as the
+    rejected ``commands_run`` claim, while still refusing to accept the masked
+    command as proof.
+    """
+    for index, message in enumerate(messages):
+        if message.tool_name != "Bash":
+            continue
+        masked_invocations: list[str] = []
+        for runtime_command in _runtime_message_command_values(message):
+            if not _has_trailing_output_filter_pipeline(runtime_command):
+                continue
+            runtime_invocation = _test_command_invocation_allowing_output_plumbing(runtime_command)
+            if runtime_invocation is not None:
+                masked_invocations.append(runtime_invocation)
+        if not masked_invocations:
+            continue
+
+        chunk = [message]
+        for following in messages[index + 1 :]:
+            if following.tool_name and not _is_tool_result_message(following):
+                break
+            chunk.append(following)
+        if not any(_message_contains_test_success(item) for item in chunk):
+            continue
+        chunk_text = "\n".join(_runtime_message_search_text(item) for item in chunk)
+        chunk_test_proof_text = "\n".join(_runtime_message_test_proof_text(item) for item in chunk)
+        if any(
+            _test_command_targets_claim(
+                command=command,
+                claim=value,
+                chunk_text=chunk_text,
+                chunk_test_proof_text=chunk_test_proof_text,
+                messages=messages,
+                task_cwd=task_cwd,
+            )
+            for command in masked_invocations
+        ):
+            return True
+    return False
+
+
 def _runtime_messages_support_file_claim(
     value: str,
     messages: tuple[AgentMessage, ...],
@@ -936,6 +1016,29 @@ def _test_command_invocation(command: str) -> str | None:
     return _test_invocation_from_shell_body(body)
 
 
+def _test_command_invocation_allowing_output_plumbing(command: str) -> str | None:
+    """Return a test invocation after stripping output plumbing unconditionally.
+
+    This must not be used as command proof. It exists only to classify rejected
+    evidence forms for diagnostics while preserving the #1208 masking guard.
+    """
+    normalized = command.strip().lower()
+    if not normalized:
+        return None
+    direct = _test_invocation_from_prefix(_strip_command_output_plumbing(normalized))
+    if direct is not None:
+        return direct
+    body = _shell_command_body(normalized)
+    if body is None:
+        return None
+    for segment, _pipefail_enabled in _segments_after_safe_shell_preamble_with_pipefail(body):
+        invocation = _test_invocation_from_prefix(_strip_command_output_plumbing(segment))
+        if invocation is not None:
+            return invocation
+        return None
+    return None
+
+
 def _shell_command_body(command: str) -> str | None:
     """Return the ``-c`` body when the command starts with a shell wrapper."""
     try:
@@ -5808,6 +5911,7 @@ Files present:
             )
 
         unsupported: list[str] = []
+        evidence_form_mismatches: list[str] = []
         backed_commands = tuple(
             command
             for command in _flatten_evidence_values(typed_evidence.get("commands_run"))
@@ -5831,6 +5935,13 @@ Files present:
                 if field_name == "commands_run":
                     if _runtime_messages_support_command_claim(value, field_messages):
                         continue
+                    if _runtime_messages_have_masked_test_command_form(
+                        value,
+                        field_messages,
+                    ):
+                        evidence_form_mismatches.append(f"{field_name}: {value}")
+                        unsupported.append(f"{field_name}: {value}")
+                        continue
                     unsupported.append(f"{field_name}: {value}")
                     continue
                 if field_name == "files_touched":
@@ -5850,16 +5961,35 @@ Files present:
                         task_cwd=self._task_cwd or self._adapter.working_directory,
                     ):
                         continue
+                    if _runtime_messages_have_masked_test_command_for_test_claim(
+                        value=value,
+                        messages=support_messages,
+                        task_cwd=self._task_cwd or self._adapter.working_directory,
+                    ):
+                        evidence_form_mismatches.append(f"{field_name}: {value}")
+                        unsupported.append(f"{field_name}: {value}")
+                        continue
                     unsupported.append(f"{field_name}: {value}")
                     continue
                 if not _runtime_messages_support_claim(value, field_messages):
                     unsupported.append(f"{field_name}: {value}")
 
         if unsupported:
+            failure_class = (
+                "EVIDENCE_FORM_MISMATCH"
+                if evidence_form_mismatches and len(evidence_form_mismatches) == len(unsupported)
+                else "FABRICATION_SUSPECTED"
+            )
+            reason_prefix = (
+                "evidence form mismatch; unprotected output-filter pipeline "
+                "cannot prove a clean command claim"
+                if failure_class == "EVIDENCE_FORM_MISMATCH"
+                else "unsupported evidence claims"
+            )
             return VerifierVerdict(
                 passed=False,
-                reasons=("unsupported evidence claims: " + "; ".join(unsupported),),
-                failure_class="FABRICATION_SUSPECTED",
+                reasons=(reason_prefix + ": " + "; ".join(unsupported),),
+                failure_class=failure_class,
             )
 
         return VerifierVerdict(passed=True)

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/src/ouroboros/orchestrator/verifier.py

@@ -85,6 +85,7 @@ class VerifierContractError(ValueError):
 _VALID_FAILURE_CLASSES: frozenset[str] = frozenset(
     {
         "EVIDENCE_MISSING",
+        "EVIDENCE_FORM_MISMATCH",
         "FABRICATION_SUSPECTED",
         "SCOPE_CREEP",
         "STALL",
@@ -105,8 +106,9 @@ class VerifierVerdict:
             to the executor — a bare FAIL with no reasons is opaque
             and indistinguishable from the first attempt).
         failure_class: Optional hint for H7 (failure taxonomy). One of
-            "EVIDENCE_MISSING", "FABRICATION_SUSPECTED", "SCOPE_CREEP",
-            "STALL", "BLOCKED", or None for an unclassified failure.
+            "EVIDENCE_MISSING", "EVIDENCE_FORM_MISMATCH",
+            "FABRICATION_SUSPECTED", "SCOPE_CREEP", "STALL",
+            "BLOCKED", or None for an unclassified failure.
     """
 
     passed: bool

{ouroboros_ai-0.40.1.dev2 → ouroboros_ai-0.40.2.dev2}/tests/unit/orchestrator/test_failure_taxonomy.py

@@ -133,6 +133,9 @@ class TestPolicyTable:
     def test_evidence_missing_retries(self) -> None:
         assert policy_for(FailureClass.EVIDENCE_MISSING).action == RecoveryAction.RETRY
 
+    def test_evidence_form_mismatch_retries(self) -> None:
+        assert policy_for(FailureClass.EVIDENCE_FORM_MISMATCH).action == RecoveryAction.RETRY
+
     def test_fabrication_escalates_model(self) -> None:
         assert (
             policy_for(FailureClass.FABRICATION_SUSPECTED).action == RecoveryAction.ESCALATE_MODEL