…dispatch (#2408)

The sync from upstream v2026.3.8 (#2360) replaced the function body with an
unconditional-throw stub, crashing every agent dispatch across Slack auto-reply,
the /agent CLI, and cron isolated-agent runs. Users saw "⚠️ Agent failed before
reply: resolveAgentRuntimeOrThrow is not available in RemoteClaw fork".

Restore the prescribed semantics: delegate to the non-throwing sibling
resolveAgentRuntime; throw only when no runtime is configured, with a message
that names the agent id and points to agents.defaults.runtime.

Return type is string (not the literal union) to avoid over-constraining
fork-native runtime names, matching resolveAgentRuntime's return shape. All 3
live callers already express this contract; no call-site changes required.

Existing call-site mocks return "claude" — they remain semantically valid.
Header comment at agent-scope.ts:437 updated from "Upstream-compat stubs
(gutted in fork)" to "Fork-native runtime & auth resolvers" since every
function in the section is fork-native.

Regression test covers both paths: configured → return, missing → throw,
with assertions on the throw-message format.

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

…dispatch (#2408) (#2423)

The sync from upstream v2026.3.8 (#2360) replaced the function body with an
unconditional-throw stub, crashing every agent dispatch across Slack auto-reply,
the /agent CLI, and cron isolated-agent runs. Users saw "⚠️ Agent failed before
reply: resolveAgentRuntimeOrThrow is not available in RemoteClaw fork".

Restore the prescribed semantics: delegate to the non-throwing sibling
resolveAgentRuntime; throw only when no runtime is configured, with a message
that names the agent id and points to agents.defaults.runtime.

Return type is string (not the literal union) to avoid over-constraining
fork-native runtime names, matching resolveAgentRuntime's return shape. All 3
live callers already express this contract; no call-site changes required.

Existing call-site mocks return "claude" — they remain semantically valid.
Header comment at agent-scope.ts:437 updated from "Upstream-compat stubs
(gutted in fork)" to "Fork-native runtime & auth resolvers" since every
function in the section is fork-native.

Regression test covers both paths: configured → return, missing → throw,
with assertions on the throw-message format.

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

Implements H9 of ADR 0005 — semantic-stub detection that H7 (throwing-stub
AST gate) and H8 (test-side mock baseline) cannot see.

Every runtime export in src/agents/ depth-1 now carries an attestation
declaring the implementation status: "live" / "stub" / "partial" /
"deprecated". The attestation-gate CI job enforces structural invariants;
CODEOWNERS enforces semantic review.

Design decisions (per user direction to proceed as issue describes):
- 4-category set including "partial" for legitimately-gutted-to-default
  functions that work in the narrow case
- Inline placement (top of file, after imports) to prevent drift
- Plain object with `as const` for tree-shakability (zero bundle cost)
- CODEOWNERS /src/agents/ → @alexey-pelykh (solo maintainer project)
- Initial scope: src/agents/* depth-1. Expand to src/middleware/, etc.,
  in follow-up issues if the pattern proves valuable

Rollout:
- 93 attested modules, 319 attestations across src/agents/* depth-1
- Auto-classified "live" by default; 11 semantic-stub exports manually
  classified "partial" (functions returning hardcoded defaults post-
  gutting — e.g., lookupContextTokens returns constant 200000,
  resolveThinkingDefault returns undefined)
- No current "stub" or "deprecated" entries (nothing currently broken)

Gate invariants:
1. Every runtime export has an attestation entry (missing fails CI)
2. Every entry corresponds to a current export (stale fails CI)
3. "live" attestation on a throwing-stub-shape function fails
4. "stub" attestation on a function with non-test callers fails

Enforced via scripts/check-attestations.mjs (AST-based classifier with
10 self-test fixtures covering all 4 invariants, expression-body arrow
enumeration, category validation, and non-function exclusion).

Polish iterations (3 total):
1. Extracted ~330 LOC of duplicated AST classification into new
   scripts/lib/throwing-stub-shape.mjs shared by check-attestations,
   generate-attestations, and check-throwing-stub-callers. Single site
   for the four calibration signals; net -373 LOC across 3 scripts,
   +210 LOC shared lib
2. Added --self-test mode with 10 fixtures + wired into CI (symmetric
   with throwing-stub-callers-gate convention). Regressions in the
   classifier itself now caught before the main gate runs
3. Aligned test-file filter across both attestation scripts via
   isTestLikeTypeScriptFile; removed 2 dead MODULE_ATTESTATIONS blocks
   from test-harness files that slipped past the naive filter

Validation (fresh-context subprocess, session 2df6a80c) initially
found one blocking issue (expression-body arrows silently excluded by
isBlock filter) + 3 non-blocking. All addressed; re-validated CLEAN.
Polish (session 9a3fc434): CLEAN after 3 iterations.

Dry-run verification:
- Probe 1 (delete attestation): FAIL "has no MODULE_ATTESTATIONS entry"
- Probe 2 (add stale attestation): FAIL "stale attestation"
- Probe 3 (live on throwing-stub): FAIL "attested 'live' but matches
  throwing-stub pattern"
- Probe 4 (stub with live callers): FAIL "has 9 non-test caller(s)"

AC 8 (PR #2360 dry-run): time-traveling to PR #2360's state is
impractical; verified the equivalent case — a "live" attestation on a
throwing-stub-shape function fails, and an unattested new export fails.
These are the precise failure modes PR #2360 exhibited.

HQ ADR 0005 H9 description updated in parallel (hq commit pending) to
reflect the as-shipped scope and 6-month sunset review.

scripts/generate-attestations.mjs is a one-shot bootstrap tool that
remains as a maintenance utility for extending the attestation scope
to new directories (idempotent; skips modules with existing blocks).

…) (#2446)

Implements H9 of ADR 0005 — semantic-stub detection that H7 (throwing-stub
AST gate) and H8 (test-side mock baseline) cannot see.

Every runtime export in src/agents/ depth-1 now carries an attestation
declaring the implementation status: "live" / "stub" / "partial" /
"deprecated". The attestation-gate CI job enforces structural invariants;
CODEOWNERS enforces semantic review.

Design decisions (per user direction to proceed as issue describes):
- 4-category set including "partial" for legitimately-gutted-to-default
  functions that work in the narrow case
- Inline placement (top of file, after imports) to prevent drift
- Plain object with `as const` for tree-shakability (zero bundle cost)
- CODEOWNERS /src/agents/ → @alexey-pelykh (solo maintainer project)
- Initial scope: src/agents/* depth-1. Expand to src/middleware/, etc.,
  in follow-up issues if the pattern proves valuable

Rollout:
- 93 attested modules, 319 attestations across src/agents/* depth-1
- Auto-classified "live" by default; 11 semantic-stub exports manually
  classified "partial" (functions returning hardcoded defaults post-
  gutting — e.g., lookupContextTokens returns constant 200000,
  resolveThinkingDefault returns undefined)
- No current "stub" or "deprecated" entries (nothing currently broken)

Gate invariants:
1. Every runtime export has an attestation entry (missing fails CI)
2. Every entry corresponds to a current export (stale fails CI)
3. "live" attestation on a throwing-stub-shape function fails
4. "stub" attestation on a function with non-test callers fails

Enforced via scripts/check-attestations.mjs (AST-based classifier with
10 self-test fixtures covering all 4 invariants, expression-body arrow
enumeration, category validation, and non-function exclusion).

Polish iterations (3 total):
1. Extracted ~330 LOC of duplicated AST classification into new
   scripts/lib/throwing-stub-shape.mjs shared by check-attestations,
   generate-attestations, and check-throwing-stub-callers. Single site
   for the four calibration signals; net -373 LOC across 3 scripts,
   +210 LOC shared lib
2. Added --self-test mode with 10 fixtures + wired into CI (symmetric
   with throwing-stub-callers-gate convention). Regressions in the
   classifier itself now caught before the main gate runs
3. Aligned test-file filter across both attestation scripts via
   isTestLikeTypeScriptFile; removed 2 dead MODULE_ATTESTATIONS blocks
   from test-harness files that slipped past the naive filter

Validation (fresh-context subprocess, session 2df6a80c) initially
found one blocking issue (expression-body arrows silently excluded by
isBlock filter) + 3 non-blocking. All addressed; re-validated CLEAN.
Polish (session 9a3fc434): CLEAN after 3 iterations.

Dry-run verification:
- Probe 1 (delete attestation): FAIL "has no MODULE_ATTESTATIONS entry"
- Probe 2 (add stale attestation): FAIL "stale attestation"
- Probe 3 (live on throwing-stub): FAIL "attested 'live' but matches
  throwing-stub pattern"
- Probe 4 (stub with live callers): FAIL "has 9 non-test caller(s)"

AC 8 (PR #2360 dry-run): time-traveling to PR #2360's state is
impractical; verified the equivalent case — a "live" attestation on a
throwing-stub-shape function fails, and an unattested new export fails.
These are the precise failure modes PR #2360 exhibited.

HQ ADR 0005 H9 description updated in parallel (hq commit pending) to
reflect the as-shipped scope and 6-month sunset review.

scripts/generate-attestations.mjs is a one-shot bootstrap tool that
remains as a maintenance utility for extending the attestation scope
to new directories (idempotent; skips modules with existing blocks).

… + rebaseline allowlist (#2355)

Restore ADR 0005 H3 (PR #2357) after it was reverted by the
v2026.3.7→v2026.3.8 sync (PR #2360, `fc2cefeeab`), which rolled
`scripts/check-no-zombie-imports.mjs` back to its pre-H3 hardcoded form
and orphaned the JSON allowlist + baselines.

Changes:

1. Restore H3 gate script with registry-derived `deadModulePatterns`,
   JSON allowlist loader, tombstone existence scan, and baseline
   comparison — identical to `af4169bf9a` minus the targeted narrowing
   in point 2.

2. Narrow `acp/runtime` fallback pattern: split broad `acp/runtime`
   into `acp/runtime/session-meta` and `acp/runtime/session-identifiers`
   (the only 4-line stubs in that namespace). `acp/runtime/errors`,
   `acp/runtime/registry`, and `acp/runtime/types` are live public API
   re-exported by `src/plugin-sdk` — the broad pattern was catching
   them as false positives.

3. Rebaseline allowlist JSON from 68 → 23 entries:
   - 42 stale entries removed (line drift during H3's offline period)
   - 11 false-positive entries removed post-narrowing
   - 8 previously-unallowed new violations retained
   - Remaining 23: 13 `agents/model-selection`, 7 `agents/sandbox`,
     3 `acp/runtime/session-meta`, 1 `agents/skills`

4. Decrement baselines:
   - `.zombie-import-allowlist-baseline`: 71 → 23
   - `.zombie-tombstone-baseline`: 12 → 0 (all tombstoned files already
     absent; #2356 is auto-satisfied)

`agents/model-selection` broad pattern is KEPT despite the file having
live H9-attested functions — those are pending gutting per the
Middleware Boundary Principle, tracked debt rather than gate false
positives. Follow-up gutting issue to be filed.

Gate passes locally: `Zombie-import gate: 36 tombstones, 45 patterns,
23 allowlisted callsites. Import scan: clean.`

Refs: ADR 0005 H3 (engineering/decisions/0005-fork-sync-hardening.md)
Regression source: #2360 (sync v2026.3.7→v2026.3.8)
Auto-satisfied: #2356 (0 tombstoned files exist)

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

… + rebaseline allowlist (#2355) (#2454)

Restore ADR 0005 H3 (PR #2357) after it was reverted by the
v2026.3.7→v2026.3.8 sync (PR #2360, `fc2cefeeab`), which rolled
`scripts/check-no-zombie-imports.mjs` back to its pre-H3 hardcoded form
and orphaned the JSON allowlist + baselines.

Changes:

1. Restore H3 gate script with registry-derived `deadModulePatterns`,
   JSON allowlist loader, tombstone existence scan, and baseline
   comparison — identical to `af4169bf9a` minus the targeted narrowing
   in point 2.

2. Narrow `acp/runtime` fallback pattern: split broad `acp/runtime`
   into `acp/runtime/session-meta` and `acp/runtime/session-identifiers`
   (the only 4-line stubs in that namespace). `acp/runtime/errors`,
   `acp/runtime/registry`, and `acp/runtime/types` are live public API
   re-exported by `src/plugin-sdk` — the broad pattern was catching
   them as false positives.

3. Rebaseline allowlist JSON from 68 → 23 entries:
   - 42 stale entries removed (line drift during H3's offline period)
   - 11 false-positive entries removed post-narrowing
   - 8 previously-unallowed new violations retained
   - Remaining 23: 13 `agents/model-selection`, 7 `agents/sandbox`,
     3 `acp/runtime/session-meta`, 1 `agents/skills`

4. Decrement baselines:
   - `.zombie-import-allowlist-baseline`: 71 → 23
   - `.zombie-tombstone-baseline`: 12 → 0 (all tombstoned files already
     absent; #2356 is auto-satisfied)

`agents/model-selection` broad pattern is KEPT despite the file having
live H9-attested functions — those are pending gutting per the
Middleware Boundary Principle, tracked debt rather than gate false
positives. Follow-up gutting issue to be filed.

Gate passes locally: `Zombie-import gate: 36 tombstones, 45 patterns,
23 allowlisted callsites. Import scan: clean.`

Refs: ADR 0005 H3 (engineering/decisions/0005-fork-sync-hardening.md)
Regression source: #2360 (sync v2026.3.7→v2026.3.8)
Auto-satisfied: #2356 (0 tombstoned files exist)

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