0.39.1 fat_harness_mode hardcoded ON without updating seed authoring guides — breaks all layered-AC seeds generated by ouroboros's own seed-architect

[buzllightyear]

Summary

In 0.39.1, the MCP execution handler hardcodes fat_harness_mode = True for all new sessions while the upstream seed authoring / QA / interview / closure components remain unchanged since 0.36. The result: seeds that follow the still-shipping seed-architect.md guidance (e.g. a layered AC pattern with a scaffold AC followed by verification ACs) fail at the very first AC because the code.yaml profile's evidence_schema.rejected_if: ["tests_passed == []"] rule rejects scaffold-only deliverables. Every downstream AC then cascade-skips with dependency_failed, and the entire run terminates with Success: 0/N.

This is a horizontal-incoherence regression: one pipeline stage's contract was tightened, the others were not, and no migration guidance / changelog warning / regression test prevented the gap from shipping.

Repro

Tested against ouroboros-ai 0.39.1 installed via uv tool install 'ouroboros-ai[mcp,claude]', Claude Code MCP host, Python 3.12 environment.

1. ooo interview <topic> — produce an interview state.
2. ooo seed — let the seed-architect agent generate the seed YAML. The agent (still using agents/seed-architect.md unchanged since 0.36) naturally produces a layered AC pattern: AC1 = scaffold (file presence + dep declaration), AC2..ACN = behavior + verification with tests.
3. ouroboros_qa passes the seed (e.g. score 0.93 / threshold 0.90). Quality bar checks internal consistency, AC measurability, ontology coverage — but not profile evidence_schema compatibility.
4. ouroboros_execute_seed (or ooo run) starts. AC1 implementation session succeeds, the worker emits [AC_COMPLETE: 1] plus a free-form JSON like {"files_touched": [...]}.
5. _fat_harness_acceptance_error rejects AC1 because typed evidence is missing commands_run / tests_passed, and rejected_if: tests_passed == [] blocks even a correctly-shaped scaffold report.
6. parallel_executor marks AC1 as failed, cascades dependency_failed to AC2..ACN, and reports Success: 0/N.

Expected vs Actual

	Expected	Actual
Authoring contract	seed-architect should be told that every AC must emit non-empty tests_passed when task_type=code, or scaffold-evidence-kind ACs should be representable in the schema	seed-architect is unchanged since 0.36, has no awareness of fat-harness requirements
QA contract	ouroboros_qa should reject seeds whose ACs cannot satisfy the active profile's evidence_schema.rejected_if rules	qa-judge dimensions are correctness / completeness / quality / intent_alignment / domain_specific — no profile_evidence_compatibility dimension
Execution contract	Accept seeds produced by the matching version's authoring pipeline	Hardcoded fat_harness_mode = True rejects layered-AC seeds even when produced by Ouroboros's own seed-architect
Release migration	Breaking change should be flagged in CHANGELOG with migration guide, or shipped as MAJOR / MINOR bump	Shipped as 0.39.0 → 0.39.1 PATCH bump, no migration guide, no deprecation period

Evidence (byte-level diff across cached versions 0.36.0 / 0.38.2 / 0.39.1)

File	0.36 → 0.39.1 change	fat-harness compatibility guidance added?
mcp/tools/execution_handlers.py:501 (fat_harness_mode = True literal)	added in 0.39.1 (0 occurrences in 0.36 + 0.38.2, 1 in 0.39.1)	n/a (this is the enforcement site)
profiles/code.yaml (evidence_schema with rejected_if: tests_passed == [])	introduced after 0.36 (file did not exist in 0.36.0 plugin cache)	n/a
agents/seed-architect.md	0 changes	❌
agents/qa-judge.md	0 changes	❌
agents/seed-closer.md	0 changes	❌
mcp/tools/subagent.py::build_generate_seed_subagent	0 changes	❌
skills/interview/SKILL.md	changed	❌ (grep "fat_harness|profile|tests_passed|rejected_if|scaffold" → 0 matches)
skills/seed/SKILL.md	changed	❌ (same grep → 0 matches; "evidence" hits are about interview/dialectical evidence, not typed evidence)
skills/run/SKILL.md	0 changes	❌

Repro of the diff (reproducible locally if any of these plugin caches are kept):

PLG=~/.claude/plugins/cache/ouroboros/ouroboros
diff -q "$PLG/0.36.0/src/ouroboros/agents/seed-architect.md" "$PLG/0.39.1/src/ouroboros/agents/seed-architect.md"  # files identical
diff -q "$PLG/0.38.2/src/ouroboros/agents/seed-architect.md" "$PLG/0.39.1/src/ouroboros/agents/seed-architect.md"  # files identical
grep -c 'fat_harness_mode = True' "$PLG/0.36.0/src/ouroboros/mcp/tools/execution_handlers.py"  # → 0
grep -c 'fat_harness_mode = True' "$PLG/0.38.2/src/ouroboros/mcp/tools/execution_handlers.py"  # → 0
grep -c 'fat_harness_mode = True' "$PLG/0.39.1/src/ouroboros/mcp/tools/execution_handlers.py"  # → 1
diff "$PLG/0.36.0/skills/seed/SKILL.md" "$PLG/0.39.1/skills/seed/SKILL.md" | grep -iE 'fat_harness|profile|tests_passed|rejected_if|scaffold'  # → 0

Asks (any one is sufficient)

1. Tighten authoring — Update agents/seed-architect.md + agents/qa-judge.md + mcp/tools/subagent.py::build_generate_seed_subagent to enforce per-AC "non-empty tests_passed is required when task_type=code" guidance. Add a profile_evidence_compatibility dimension to qa-judge.
2. Relax fat-harness — Allow scaffold-evidence-kind ACs in _fat_harness_acceptance_error: if tests_passed == [] AND the AC text matches a scaffold heuristic (no "test"/"assert"/"verify" keyword AND files_touched != []), accept. Or add an explicit evidence_kind: scaffold | verification field to the seed AC schema.
3. Revert default — Make fat_harness_mode opt-in via seed.orchestrator.execution_mode: fat_harness (the inverse of the deleted legacy opt-in), and default to False until the authoring pipeline can produce fat-harness-compatible seeds reliably. Document the migration in CHANGELOG.

The minimal acceptable fix is #3 (single-file revert with documentation), but #1 is the structurally correct solution that closes the loop between authoring and execution.

Workaround

Two-line patch in the installed package:

• mcp/tools/execution_handlers.py:501 — fat_harness_mode = True → fat_harness_mode = False
• cli/commands/run.py::_resolve_fat_harness_mode — return True → return False

Survives runtime but is wiped by every uv tool upgrade ouroboros-ai. Also patches ~/.claude/plugins/cache/ouroboros/ouroboros/0.39.1/.mcp.json to add --python 3.12 to the uvx args (separate, unrelated issue: uvx picks system Python 3.10 by default which fails to resolve the >=3.12 requirement).

Environment

• ouroboros-ai 0.39.1 (uv tool install)
• Claude Code MCP host
• macOS Darwin 25.5.0 / Python 3.13 (uvx pinned to 3.12 for ouroboros)
• Affected project: 4 prior phases (3.1, 3.2, 3.3, 3.4) of a brownfield FastAPI/RN monorepo successfully executed on 0.36/0.38 with the same seed-architect output style; first phase attempted on 0.39.1 (4.1) fails reproducibly on AC1.

[Q00]

Opened PR #1203 with the minimal safe mitigation: fresh CLI/MCP seed execution now uses fat-harness acceptance only when seed.orchestrator.execution_mode: fat_harness is explicitly set. Targeted tests passed.

Posted by agentos-roadmap-warden — bot. Reply with /warden ignore to suppress further comments on this thread.

[ouroboros-agent]

There is already an open PR addressing this regression: #1203.

The issue is well-scoped as a bug with a detailed reproduction, and the linked PR appears to implement the minimal safe mitigation described in the report.

ouroboros-agent[bot]