Skip to content
PETROVA

SP-1.5 seam-defect — agent:-prefix mismatch breaks the C2↔I1 join

REMEDIATED — GREEN (2026-05-18). Plan owner chose Option A (probe/prompt-side; principal := agent:petrova-cairnet-bridge). Corrective round CR1–CR3: CR1 (pebble d6b244e) added the missing cross-repo POST→GET seam test and empirically proved the recorded value is agent:petrova-cairnet-bridge (real auth + cairn POST + Postgres; bare form matches zero stones); CR2 (eva-hq 7fe44dc) set the prompt producer-principal value to the prefixed form (owner/producer split preserved); CR3 (petrova-codes 6110836) realigned fixtures + added a contract-pin test locking the prefixed literal and its %3A URL encoding. The terminal whole-seam gate round 2 returned GREEN: SC-1 delivered end-to-end, C1/C2/C3/I1/I2/I3 each closed, the agent:-prefix join and the %3A round-trip verified from source, CR1–CR3 spec-clean, regression clean (only the 4 known pre-existing failures, base identity proven). Nothing merged at time of this note; all work on the four feature branches awaiting branch-finishing.

Tracked next-round residuals (MR-2 — next round’s input, NOT bolted onto this GREEN round):

  1. eva-hq prompts/petrova-wire-rocky/body.md lines ~72 & ~328: the human-facing sc-2 verification curl example uses ?agent_id=/&since=, but Pebble’s stones-GET accepts only ?agent= (no ?since=; cairn.py:250). A literal sc-2 curl would hit the unknown-param-ignored path → unfiltered listing → spurious pass: a lying verification example (same defect class as the original SP-1 seam-defect, but in a diagnostic example, not the load-bearing probe — rocky.ts:91 correctly uses ?agent=). Doc-correctness fix: align the sc-2 example to ?agent= and drop the &since= (probe windows client-side).
  2. airlock scripts/seed-petrova-cairnet-bridge.ts header doc-comment lists only cairn:emit+knowledge:register; the executable buildBridgeClientSpec() correctly returns all three incl. cairn:read. Doc-comment nit; executable spec is correct and unit-tested.

SP-1.5 seam-defect — agent:-prefix mismatch breaks the C2↔I1 join

Section titled “SP-1.5 seam-defect — agent:-prefix mismatch breaks the C2↔I1 join”

Disposition

Section titled “Disposition”

All five SP-1.5 tasks (R1–R5) individually passed implement → spec-review → code-quality-review (R3/R4/R5 through multiple review rounds that caught and closed real per-task defects). The mandatory terminal whole-seam gate then found that the assembled cross-repo seam still does not deliver SC-1 end-to-end, because of a join no single task could see.

Verdict: SEAM-DEFECT-FOUND. Nothing is merged; all work is on the four feature branches (feat/cairn-idempotency, feat/petrova-cairnet-bridge-client, feat/cairnet-seam-sp1, feat/petrova-wire-rocky-deterministic); no repo’s main is affected. Branch-finishing was not invoked and SC-1 success is not claimed. Per MR-2 and the plan’s terminal clause this is the next round’s input, not an R6 bolt-on.

The defect — the C2↔I1 load-bearing join

Section titled “The defect — the C2↔I1 load-bearing join”

The SP-1.5 binding decision (principal := the OAuth client_id petrova-cairnet-bridge, no -001) was applied coherently across R1/R2/R3/R4/R5. But it rested on an unverified assumption: that Pebble records an emitted stone’s agent_id as the bare introspected principal. It does not.

Source trace (re-derived from primary source at the terminal gate, not from task reports):

  1. airlock R2 — scripts/seed-petrova-cairnet-bridge.ts:55: clientId: "petrova-cairnet-bridge", scopes ["cairn:emit","cairn:read","knowledge:register"]. Bare client_id.
  2. Pebble R1 — src/pebble/api/middleware/auth.py Bearer block: sub = body.get("sub") or body.get("client_id") → for a client_credentials token, sub = "petrova-cairnet-bridge"; metadata["agent_id"] = "petrova-cairnet-bridge". Correct.
  3. Pebble — the break — src/pebble/api/routes/cairn.py:148-154 _principal_triple(): binding = user.metadata.get("agent_id") → return (f"agent:{binding}", ...). The POST /stones handler (cairn.py:321,339) persists the stone with agent_id = "agent:petrova-cairnet-bridge" (the request-body envelope agent_id is correctly ignored — cairn.py:312 — so the petrova-codes envelope value is moot; identity is principal-derived and gets the agent: prefix).
  4. Pebble stones-GET filter — src/pebble/infrastructure/services/cairn_service.py:174-175: if agent: stmt = stmt.where(CairnStoneModel.agent_id == agent) — exact equality, no prefix normalization.
  5. petrova-codes R3 probe — cli/src/probes/rocky.ts:89: ?agent=${ev.cairnet_agent_id} where ev.cairnet_agent_id = petrova-cairnet-bridge (bare, per the eva-hq R5 prompt’s <<CAIRNET_PRODUCER_PRINCIPAL>>).

Result: stone stored under agent:petrova-cairnet-bridge; probe asks ?agent=petrova-cairnet-bridge; WHERE agent_id == 'petrova-cairnet-bridge' returns 0 rows. The rocky probe always reports total = 0 → failing on a perfectly-wired repo. SC-1 link 3 (and transitively the §3 join, the effective behavior of I1/I2, sc-2/sc-3) breaks.

Why no task caught it: R1’s bearer test asserts metadata['agent_id'] at the middleware boundary (fixture sub: "stratt-hq-bridge-001") and never crosses _principal_triple; R3 tests mock Pebble responses; R5 is a prompt. The agent: prefix is introduced entirely inside Pebble’s cairn.py and is asserted by no cross-repo test.

Per-defect status (assembled)

Section titled “Per-defect status (assembled)”
  • C1 / C3 (Pebble Bearer auth): CLOSED. RFC7662, basic-auth, fail-closed, issuer-checked, JWT-shape-gated; session/apikey paths untouched; client_credentials resolves agent_id from client_id.
  • C1↔R2 join: CLOSED. R1 parses scope→permissions+scopes; R2 supplies cairn:emit+cairn:read; both require_permission gates pass; R1 does not depend on airlock-custom metadata ({} is fine — agent_id derives from sub/client_id).
  • C2 (intrinsic emit, R4): CLOSED as written — unconditional rocky emit after all refusal gates, deterministic emitted_at→ first_emitted_at, ANCHOR_REWRITE_INVARIANT (no silent drop), queued path still anchors, posts Bearer to the auth-gated /api/cairn route (connects to R1’s path).
  • ★ C2↔I1 (the §3 join): NOT CLOSED — the agent:-prefix mismatch above. THE defect.
  • I1 / I2 (probe↔Pebble contract): contract-shape CLOSED (?agent= is the real param — cairn.py:250; stoneType/ createdAt are the real _stone_to_dict keys — cairn_service.py:71-77; no ?since=; client-side createdAt window) but functionally void because the ?agent= value can never match a recorded stone.
  • I3 / R5: prompt internally consistent and matches R3’s ?agent= param name, but instructs the bare principal Pebble does not record. R5 faithfully implements the settled contract; the settled contract is wrong about what Pebble records.

Corrective direction (next round input — pick ONE canonical end)

Section titled “Corrective direction (next round input — pick ONE canonical end)”

Do not patch in two places. Two options:

  • A — probe/prompt/evidence side (recommended; smallest, no code change, preserves Pebble’s agent/human namespacing doctrine). Change <<CAIRNET_PRODUCER_PRINCIPAL>> in eva-hq/prompts/petrova-wire-rocky/body.md (rocky evidence cairnet_agent_id, stones-GET examples, sc-2 curl) to agent:petrova-cairnet-bridge. petrova-codes R3 passes ev.cairnet_agent_id through verbatim — no code change. Add one cross-repo integration test: POST a stone as the bridge principal, then GET ?agent= with the documented value, assert ≥1 match.
  • B — Pebble side. Strip/skip the agent: prefix for machine bridge principals in _principal_triple so recorded agent_id == "petrova-cairnet-bridge". Larger blast radius — touches Pebble’s identity-provenance doctrine for all agent writes. Not recommended unless the doctrine explicitly wants bridge clients un-prefixed.

The absence of a single POST-then-GET-by-?agent= cross-repo test (with the exact documented value) is the structural root cause of this slipping past five green tasks; that test is mandatory in the next round whichever option is chosen.

Residual observations (non-blocking, carried forward)

Section titled “Residual observations (non-blocking, carried forward)”
  • Pebble _introspect_opaque sync httpx.post blocks the async event loop — fail-closed, faithful KAHN port; deferred performance fast-follow, not a seam defect, not an SC-1 blocker.
  • cairnet_agent_id field-name vs value — benign-but-misleading naming wart; becomes more confusing once the value must carry an agent: prefix. The next round’s decision doc should call it out.
  • 4 pre-existing petrova-codes cli failures — tests/registry.test.ts ×3 (registry-applicability data) + tests/validate.test.ts ×1 (MR-4 filename validator). Orthogonal to this seam; tsc clean; 309–313 passing depending on round. The MR-4 violator is the pre-existing docs/decisions/2026-05-15-petrova_act_registry_edit-choco-hq.md (underscores in slug) — out of SP-1.5 scope; do not rename it here (would compound the round). Hyphen-only new doc slugs pass.

Verification status of these findings

Section titled “Verification status of these findings”

Each source line above was read at the terminal gate from primary source in the respective repo at the stated feature-branch HEADs (pebble 108b33f, airlock 59da1bee, petrova-codes feat/cairnet-seam-sp1 HEAD incl. 60375f2/3718ebb, eva-hq e2a94bf). The full petrova-codes cli suite was re-run at the gate: 4 failed / 313 passed, tsc --noEmit clean — the 4 failures are the pre-existing orthogonal set above.

SP-1.6 resolution (2026-05-18) — both tracked residuals closed

Section titled “SP-1.6 resolution (2026-05-18) — both tracked residuals closed”

Issue #126’s two next-round residuals are remediated by SP-1.6 (spec docs/specs/2026-05-18-sp1.6-stones-get-strict-params-design.md, plan docs/plans/2026-05-18-sp1.6-stones-get-strict-params-implementation.md, caller-compat audit docs/findings/2026-05-18-sp1.6-stones-get-caller-audit.md). Each task ran implement → spec-review → code-quality-review.

  • Residual 1 (the systemic fail-open, not just the sc-2 instance). Investigation reframed it: Pebble’s GET /api/cairn/stones silently ignored unknown query params → unfiltered listing → any mistyped/legacy verification could falsely pass. T2 (pebble feat/sp1.6-stones-get-strict-params 4657a535 + review-hardening 823f100) makes the endpoint reject unknown params with 400 (fail-closed), naming offenders + the allowed set derived from the single source-of-truth constant; a pure-unit allowlist/​message-rot drift guard runs DB-free (not CI-only). T1 (caller-compat audit 679958303 + accuracy correction 363d884f) proved no first-party caller sends a non-allowlisted param before the 400 was enabled (gate cleared). T3 (eva-hq feat/sp1.6-sc2-curl-fix bd34af8) corrects the human sc-2 verification curl to ?agent=agent%3Apetrova-cairnet-bridge with no since (Pebble has none; probe windows client-side) — the lying example can no longer pass vacuously, and post-T2 the old form 400s loudly. SP-1.5 R5/CR2 owner/producer split preserved.
  • Residual 2. T4 (airlock feat/sp1.6-seed-comment-fix 078333f63) makes the seed file header doc-comment enumerate cairn:read, matching the executable buildBridgeClientSpec(). Comment-only.

State: all four tasks GREEN through both review stages; work is on the four feat/sp1.6-* feature branches, NOT yet merged (no repo main affected). Issue #126 stays open until these branches are integrated — closed referencing the merge, per the same don’t-claim-done-before-integration discipline applied to SP-1.5.