feat(auto): wire driver selection through capabilities (#672)

* Add selected driver and brake mode for auto interviews

* fix(auto): align selected driver resume state

* feat(auto): wire driver selection through capabilities

* fix(auto): prompt for interview driver selection

* fix(auto): only prompt installed interview drivers

* fix(auto): preserve driver scaffold ledger sources

* fix(auto): scope add-keyword brake gate to feature additions

The bot's design note on PR #672 flagged a regression: bare `add` in the
risk gate matched ordinary CRUD wording like "How should users add a
task?", false-blocking selected-driver sessions under brake=on.

Replace bare `add` with `add(?:ing)?\s+(?:a|an|the)?\s*<scope-noun>`
where the scope-noun is feature/capability/support/requirement/epic/
story/scope/product-area. CRUD verbs no longer match while genuine
scope-additions ("Should we add a feature for offline mode?") still
fire the gate.

Refs #672

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(auto): align selected-driver ledger values with the freeform answer

The bot's blocking finding on PR #672 noted that the selected-driver
path sent the driver's freeform answer into the interview transcript
but kept the deterministic scaffold's value in ``ledger_updates``.
Seed generation reads the transcript while grading reads the ledger,
so the two could disagree silently (driver says "Django", ledger says
"FastAPI") without anyone catching the divergence.

- Replace each scaffold-derived ledger entry value with the driver's
  freeform answer verbatim and tag the entry CONFLICTING with reduced
  confidence (<= 0.4), so the Seed-ready/A-grade gates surface the
  divergence instead of treating the scaffold value as confirmed.
- Preserve the structural section/key/source so downstream
  section-aware Seed generation still works, and keep the original
  scaffold value in the rationale + evidence as audit context.
- Update the existing ledger-values test to lock in the new contract
  and assert the original scaffold value is preserved as rationale
  evidence.

Refs #675

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* test(auto): lock packaged driver/brake forwarding at MCP payload layer (#679)

Issue #673 acceptance criteria require that the actual MCP payload reaching
ouroboros_auto carries forwarded driver/brake values and never leaks $driver
or $brake placeholder literals. Upstream stack/3 already locks dispatch-level
resolution; this commit adds the missing layer by mocking the ouroboros MCP
server and asserting call_tool's payload.

- Forwarding test: --driver hermes --brake off must reach call_tool with
  driver="hermes", brake="off", skip_run=True.
- Placeholder leak test: plain `ooo auto "goal"` must not send "$driver" or
  "$brake" as real values; an empty/None value is acceptable.

Closes #673

* fix(auto): use INFERRED status for driver-derived ledger entries

The earlier fix that aligned ledger values with the driver's freeform
answer marked every structural ledger update as ``LedgerStatus.CONFLICTING``.
The interview loop treats CONFLICTING (along with MISSING/WEAK/BLOCKED) as
an open gap (``ledger.py:open_gaps``), so a selected-driver session could
populate every required section and still never reach ``is_seed_ready`` —
the loop would carry the same gaps forever and eventually block.

Use ``LedgerStatus.INFERRED`` instead. The ledger entry still:
- Carries the driver's freeform answer verbatim as the value, so the
  persisted ledger and the interview transcript share the same content
  (no silent divergence between the two sources of truth).
- Has reduced confidence (<= 0.4) and an ``auto_interview_transcript``
  evidence marker so grading and the A-grade gate can downgrade or
  re-verify the driver answer before final acceptance.
- Preserves the original scaffold value verbatim in the rationale and
  in a ``scaffold_value:<value>`` evidence tag, so divergence between
  scaffold and driver is never silently lost.

INFERRED is the right semantic: the value is the driver's inferred answer,
not the user's confirmed choice; downstream grading is responsible for
deciding whether to escalate to verification before A-grade. The interview
loop is now free to converge for selected-driver sessions.

Tests:
- ``test_driver_answerer_ledger_values_reflect_driver_answer_without_blocking_loop``
  asserts the new INFERRED+verbatim+scaffold-evidence contract and locks
  in that the status is not in the loop's blocking set.
- ``test_driver_answerer_ledger_does_not_block_seed_ready_convergence``
  is a regression test that drives the answerer through every required
  section and asserts those sections do not end up in the open-gap set
  (the contract that broke under the CONFLICTING variant).

Addresses ouroboros-agent[bot] blocking finding on PR #672 against
``f229e35``. Refs #672.

* feat(auto): add selected-driver answer metadata (#682)

Co-authored-by: Hermes Agent <hermes-agent@users.noreply.github.com>
Co-authored-by: shaun0927 <junghwan1912@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(auto): align mcp resume bounds (#680)

Co-authored-by: Hermes Agent <hermes-agent@users.noreply.github.com>

---------

Co-authored-by: shaun0927 <junghwan1912@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Junghwan <70629228+shaun0927@users.noreply.github.com>
Co-authored-by: Hermes Agent <hermes-agent@users.noreply.github.com>

Author: 5 people

skills/auto/SKILL.md

@@ -9,6 +9,8 @@ mcp_args:
   max_interview_rounds: "$max_interview_rounds"
   max_repair_rounds: "$max_repair_rounds"
   skip_run: "$skip_run"
+  driver: "$driver"
+  brake: "$brake"
 ---
 
 # /ouroboros:auto
@@ -30,6 +32,7 @@ is unavailable. A manual fallback is not an `ooo auto` run.
 ooo auto "Build a local-first habit tracker CLI"
 ooo auto --resume auto_abc123
 ooo auto "Build a local-first habit tracker CLI" --skip-run
+ooo auto "Build a local-first habit tracker CLI" --driver hermes --brake on
 /ouroboros:auto "Build a local-first habit tracker CLI"
 ```
 
@@ -42,3 +45,10 @@ ooo auto "Build a local-first habit tracker CLI" --skip-run
 5. Starts execution only after A-grade.
 
 The pipeline must not hang indefinitely: all loops are bounded and timeout failures return a resumable `auto_session_id`. Resume with `ooo auto --resume <auto_session_id>`. Use `--skip-run` to stop after the A-grade Seed. The CLI-only `--show-ledger` flag prints assumptions/non-goals; MCP skill responses already include the same ledger summary when available.
+
+When invoked through the interactive CLI without `--driver` or a configured
+default driver, `ooo auto` asks whether to use a selected interview driver if
+one of the supported driver CLIs is installed. Declining that prompt, or having
+no installed driver CLI, keeps the deterministic auto answerer. Use
+`--driver <backend>` to select a driver explicitly; use `--brake on|off` to
+control whether risky driver answers block for approval.

src/ouroboros/auto/__init__.py

@@ -6,7 +6,7 @@
 before starting execution.
 """
 
-from ouroboros.auto.answerer import AutoAnswer, AutoAnswerer, AutoAnswerSource
+from ouroboros.auto.answerer import AutoAnswer, AutoAnswerer, AutoAnswerMetadata, AutoAnswerSource
 from ouroboros.auto.grading import GradeGate, GradeResult, SeedGrade
 from ouroboros.auto.interview_driver import AutoInterviewDriver, AutoInterviewResult, InterviewTurn
 from ouroboros.auto.ledger import LedgerEntry, LedgerSection, SeedDraftLedger
@@ -17,6 +17,7 @@
 
 __all__ = [
     "AutoAnswer",
+    "AutoAnswerMetadata",
     "AutoAnswerSource",
     "AutoAnswerer",
     "AutoInterviewDriver",

src/ouroboros/auto/answerer.py

@@ -18,6 +18,7 @@ class AutoAnswerSource(StrEnum):
     EXISTING_CONVENTION = "existing_convention"
     CONSERVATIVE_DEFAULT = "conservative_default"
     ASSUMPTION = "assumption"
+    DRIVER = "driver"
     NON_GOAL = "non_goal"
     BLOCKER = "blocker"
 
@@ -64,6 +65,15 @@ class AutoBlocker:
     question: str
 
 
+@dataclass(frozen=True, slots=True)
+class AutoAnswerMetadata:
+    """Structured provenance for auto answers that need audit context."""
+
+    risk: str | None = None
+    confidence: float | None = None
+    provenance: tuple[str, ...] = ()
+
+
 @dataclass(frozen=True, slots=True)
 class AutoAnswer:
     """Answer plus structured ledger updates."""
@@ -75,6 +85,7 @@ class AutoAnswer:
     assumptions: list[str] = field(default_factory=list)
     non_goals: list[str] = field(default_factory=list)
     blocker: AutoBlocker | None = None
+    metadata: AutoAnswerMetadata = field(default_factory=AutoAnswerMetadata)
 
     @property
     def prefixed_text(self) -> str: