# Core Architecture v2 S5 Delegation Assessment

**Slice:** S5 — Progressive Context & Paseo Session Identity
**Checkpoint:** `1f92501c0f63199500b99b07ab202adb283d6ef1`
**Branch:** `core-architecture-v2`
**Authority:** `docs/CORE_ARCHITECTURE_V2.md` §6–7 is frozen normative target. This tactical graph records S5 implementation ownership only; it does not amend that target.

## Entry checks and dependency decision

- Initial checkout was clean on `core-architecture-v2`; `HEAD` exactly matched the accepted S3 checkpoint above.
- S1 execution identity is accepted. `ExecutionBindingV2` carries operation execution revision, participant/generation, candidate, controller epoch, blueprint/policy, actual session, and actual `ContextManifest` / `PromptManifest` digests. Launch adapters and StructuredResult validation compare the complete binding.
- S3 authority is accepted. `assertCurrentControllerOwner` requires the current durable epoch and controller token; `assertControllerEpoch`, current `OperationRecord`, frozen policy, candidate, participant registration, and `ExecutionAuthorityV1` / capability leases provide deterministic inputs for a controller-issued context grant.
- The existing `ContextRefAuthorizationV1` is only a self-digested data shape: operation/session/candidate are optional, participant is not compared to an invocation identity, and a public pure helper can mint it. `ContextRefV1` is consumed only in unit tests. Existing retrieval has a separate logical-agent allowlist and does not persist candidate/session-bound receipts or enforce current S3 authority.
- `PaseoSessionBindingV1` is consumed only by its unit tests. It currently omits candidate, policy, context/prompt digests, and controller epoch from reuse equality; production worker session reuse does not call it.
- **Dependency decision:** the shared interface can be implemented without changing §7 or accepted S3 semantics. S5 will not issue grants from model/participant output. The controller issuer will require the current controller fence and current operation state, a registered participant, exact candidate/policy/blueprint/epoch identity, current read authority, and the actual materialized session. Retrieval will revalidate its durable issue record against the current bound execution identity before returning bytes.

## Frozen S5 ContextRef authorization contract

`ContextRefAuthorizationV1` is a controller-issued, expiring, participant-execution-bound capability. The controller issue path derives it from current durable state; neither a model nor participant can submit or mint authority. Its canonical identity binds:

- operation ID and execution revision;
- project ID and exact candidate revision/digest;
- participant ID and generation;
- ExecutionBlueprint digest, frozen policy digest, and controller epoch;
- the actual Paseo session ID;
- the exact allowed ContextRef IDs, shard IDs, locators, artifact paths, content digests, and source digests;
- issue and expiry instants, plus the canonical grant digest.

The grant digest proves payload integrity only. Controller origin is established by the controller-only issuer and durable issue record. Once S1 binds the final manifest and prompt, the durable issue record also pins the complete `ExecutionBindingV2` digest. At each retrieval the deterministic gateway checks the current operation/candidate/policy/epoch/participant binding, session, grant expiry, exact permitted ref, source receipt and bytes, request replay identity, and remaining operation/turn budget. It emits an immutable receipt bound to the same candidate/session/execution binding and delivered-byte digest. Any mismatch, stale generation, expired grant, changed source, unknown ref, or replay is rejected.

Continuation records are controller-persisted and bind the source and destination session/turn, sequence, candidate, participant generation, execution binding, prompt/context manifests, referenced refs and retrieval receipts, and controller epoch. Loading or advancing a continuation revalidates all bindings and rejects stale/replayed records. A changed prompt or context requires a new execution binding as required by §7; S1's existing binding remains authoritative for launch and StructuredResult acceptance.

## Tactical WorkGraph

Mechanisms name the meaningful decision in each unit. Deterministic checks remain authoritative for identity, source bytes, grants, receipts, budgets, and lifecycle. No WorkUnit enters S4, S6, S7, S8, S9, or another slice.

| ID / owner | Mechanism | Dependencies | Files / areas | Invariants | Acceptance criteria and exact tests |
| --- | --- | --- | --- | --- | --- |
| S5-W0 — LUNA_OWNED contract freeze and assessment | HYBRID: architectural interpretation constrained by deterministic S1/S3 contracts | Accepted S1 and S3; completed before implementation | This file; `src/architecture/executionIdentity.ts`; `src/security/executionLease.ts`; `src/operations/state.ts`; `src/context/runtimeV2.ts`; `src/paseo/sessionBinding.ts` | §7 remains unchanged; context cannot grant capability, mutation, WorkGraph ownership, or authority; no model-issued authorization | Confirm current owner/epoch/candidate/policy/participant/lease fields; confirm exact S1 manifest/result binding. Inspected baseline: `tests/contextRuntimeV2.test.ts`, `tests/agentPromptStructuredResultProvenance.test.ts`, `tests/structuredResultGateway.test.ts`, `tests/paseo/sessionBinding.test.ts`, `tests/paseoPendingExecutionIdentity.test.ts`, `tests/paseoInitialTurnBarrier.test.ts`. |
| S5-W1 — FLASH_DELEGATABLE full canonical PaseoSessionBindingV1 contract | DETERMINISTIC | W0; isolated Paseo worktree; exact binding fields frozen above and §7 | `src/paseo/sessionBinding.ts`; `tests/paseo/sessionBinding.test.ts` | Strict full identity comparison includes operation, participant/generation, candidate, blueprint, policy, context/prompt, and controller epoch. Missing expected identity is invalid; no wildcard or legacy read. Every mismatch is non-reusable and follows explicit rotation/new-session path. | Extend V1 binding and durable validation; create/load/idempotent bind only for complete canonical identity; concurrency-safe rotation increments session generation; ACTIVE only; corrupted/stale/archived/lost/mismatched binding cannot be reused. Run `npx vitest run tests/paseo/sessionBinding.test.ts`; report exact result. No child agents. |
| S5-W2 — LUNA_OWNED controller-issued ContextRef authorization | DETERMINISTIC | W0; W1 contract merged/reviewed before production session integration | `src/context/runtimeV2.ts`; new controller authorization module under `src/context/`; `src/operations/state.ts` only if lifecycle persistence requires it; `src/security/executionLease.ts`; context authorization tests | Only current controller owner issues; validate current durable operation, policy, candidate, epoch, registered participant, generation, current read lease, blueprint, actual session, exact allowed ref/source digests, and expiry. Public/pure digest creation never constitutes authority. No source edits authorized by a context grant. | Reject forged/self-issued, expired, wrong operation/participant/generation/candidate/blueprint/policy/epoch/session/fragment grants, source/receipt tampering, and stale controller state. Accept only controller-issued current exact binding. Add focused tests in `tests/contextRuntimeV2.test.ts` and dedicated authorization tests. |
| S5-W3 — LUNA_OWNED durable retrieval and continuation lifecycle | DETERMINISTIC | W2; receipt and binding schemas fixed before integration | `src/context/retrieval/`; `src/context/runtimeV2.ts`; new continuation persistence module under `src/context/`; `tests/contextRuntimeV2.test.ts`; new retrieval lifecycle tests | Retrieval and continuation are durable, operation/candidate/participant-generation/session/binding/epoch-bound; source and delivered digests are checked; unique request IDs prevent receipt replay; retrieval limits and usage are durably accounted per exact execution binding/session; stale context cannot resume. | Exact authorization is required for retrieval; source bytes must match the authorized receipt; budget overflow, path/symlink escape, expired/stale identity, wrong session, continuation sequence/ref mismatch, and replay fail closed. Verify receipt content/digests and current-binding checks. Add focused retrieval/continuation tests. |
| S5-W4 — LUNA_OWNED progressive role projection and manifest/prompt boundary | HYBRID: optional projectable-fragment compression is bounded by deterministic eligibility/budget/provenance; normative/authority text is deterministic verbatim | W2–W3; exact grant/ref schema; accepted S1 prompt/result invariants | `src/context/gateway.ts`; `src/context/projectors/`; `src/workers/agentPrompt.ts`; `src/paseo/launchSpec.ts`; relevant result provenance adapter | Broader role projection may summarize eligible projectable fragments only. TaskContracts, sealed requirements, authority envelopes, schemas, policy, assertions, and accepted skill procedures remain exact. The immutable S1 launch ContextManifest/PromptManifest digests cover the exact prompt, addressable ref/source identities, and retrieval budget; S5 adds result-bound ProgressiveContextManifestV1 evidence over the exact retrieved receipt set, delivered-byte digests, durable usage/state, and those launch digests. A changed launch context/prompt changes ExecutionBinding. | Exercise the assignment-scoped context composition and existing kind-based projectors; verify Planner/Explorer structured handoffs use their existing role projectors, other handoffs retain source fallback, non-handoff kinds retain their existing projectors, normative bytes and skill procedures remain verbatim, and actual prompt/context digests survive launch/result adapters. Run `tests/contextArchitectureClose.test.ts`, `tests/contextEfficiency.test.ts`, `tests/agentPromptStructuredResultProvenance.test.ts`, `tests/structuredResultGateway.test.ts`, and affected adapter suites. |
| S5-W4a — FLASH_DELEGATABLE deterministic Planner/Explorer handoff routing | DETERMINISTIC | W0; existing `projectPlanner` / `projectExplorer`; isolated Paseo worktree | `src/context/gateway.ts`; `tests/contextEfficiency.test.ts` | Only structured handoff fragments route through the existing deterministic role projectors; all other fragment-kind dispatch, verbatim preservation, and fallback projection stay unchanged. | Verify Planner and Explorer routing from the effective recipient role, agent-scoped role fallback, non-role handoff fallback, normative bytes, and kind-based non-handoff projections. Run `npx vitest run tests/contextEfficiency.test.ts` and `npx tsc -p tsconfig.json --noEmit`. Worker reported 17/17 tests and typecheck exit 0; integrated main checkout also passes in the 12-file / 118-test S5 focused lane. No child agents. |
| S5-W5 — LUNA_OWNED Paseo session create/reuse/rotation integration | DETERMINISTIC | W1–W4; S1 actual-session launch boundary | `src/paseo/sessionBinding.ts`; `src/paseo/launchSpec.ts`; `src/paseo/runtimeCore.ts`; `src/paseo/sdk.ts`; `src/workers/agentPrompt.ts`; launch/result adapter tests | Reuse only on exact canonical binding equality. A mismatch or missing/corrupt binding rotates or starts a new session before prompt dispatch; result provenance names the actual returned session. Never relabel an old live session. | Test exact reuse and one-at-a-time mismatch cases for operation, participant, generation, candidate, blueprint, policy, context, prompt, and epoch; rotation occurs before first prompt; returned/session-bound result IDs agree. Run `tests/paseo/sessionBinding.test.ts`, `tests/paseoPendingExecutionIdentity.test.ts`, `tests/paseoInitialTurnBarrier.test.ts`, and `tests/agentPromptStructuredResultProvenance.test.ts`. |
| S5-W6 — LUNA_OWNED integration, review, evidence, and living status | HYBRID: semantic final acceptance plus deterministic test/build evidence | W1–W5 | S5 affected rows in `docs/CORE_ARCHITECTURE_V2_STATUS.md`; `docs/CORE_ARCHITECTURE_V2_CONFORMANCE.md`; `docs/ENGINEERING_LEDGER.md` only for evidence-backed closure/findings; affected source/tests | Preserve accepted S1/S2/S3 and unrelated work; do not edit frozen target; do not start another slice; do not commit/push. | Run focused S5/adapter suites, `npx tsc -p tsconfig.json --noEmit`, `npm run build`, affected Vitest suites, full non-system Vitest, and only relevant S5 deterministic system suites. Record exact exit/counts and keep REAL_PROVIDER/BROWSER/OPA/ADVERSARIAL/PACKED_E2E claims separate. Re-read diff, status, and frozen docs. |

## Delegation assessment

Two safely separable meaningful implementation units are delegated. S5-W1 replaces the unused partial session-binding contract and its focused tests. Agent `3023b545-8fd0-4f7d-a83e-84a2acf11c3d` used worktree `/home/james/.paseo/worktrees/0jx5pvzi/core-v2-s5-session-binding` (`s5-session-binding-flash`). S5-W4a routes only structured Planner/Explorer handoffs through existing deterministic projectors and tests that narrow behavior. Agent `c63e1e54-6a30-4326-80c7-238387317a81` used worktree `/home/james/.paseo/worktrees/0jx5pvzi/core-v2-s5-progressive-projection` (`s5-progressive-projection-flash`). Neither unit decides authority, context authorization, manifest semantics, or production session policy. The projection dispatch uses the effective recipient role because production fragments have no producer-role marker; no new metadata convention was introduced. Worker-reported results: W1 session binding 17/17 tests and typecheck exit 0; W4a projection 17/17 tests and typecheck exit 0. The recovery test double was updated to provide the newly required durable state-root resolver. Those remain sequential with LUNA because they cross S1 binding, S3 controller authority, retrieval receipts, prompt construction, and StructuredResult provenance.

Configured profile copied from Paseo `list_profiles` for both delegated units: **DeepSeek V4.1 Flash Implementer** — provider `opencode`; model `opencode-go/deepseek-v4.1-flash`; `settings.modeId = "build"`; `settings.thinkingOptionId = "max"`; `settings.features = { "auto_accept": true }`. Launches are restricted to separate Paseo-managed worktrees. S3–S5 Flash workers must not create child agents.

No `NO_DELEGATION_JUSTIFICATION`: Flash is used for the bounded canonical binding schema and tests. The coupled authority, controller issuance, prompt/context, and result integration units are intentionally LUNA-owned.

## Evidence and state

At the pre-remediation S5 checkpoint, the prior lead reported focused S5/adapter tests at 12 files / 118 tests; TypeScript and build passed; full non-system Vitest passed 320 files with 830 passed and 7 pending; and the three relevant deterministic system suites passed 45 tests. That historical gate evidence is superseded for current status by the remediation evidence below. The frozen target remained untouched and no commit or push was made at that checkpoint.

## S5 same-slice remediation assessment — 2026-09-24

### Bounded finding and frozen contract

The S5 lead report exposed a manifest-to-grant gap in `src/workers/agentPrompt.ts` and `src/context/authorizationV2.ts`: launch binding checked the canonical S1 `ContextManifest` digest, but controller grant issuance independently reread the durable `ContextEnvelope` and authorized its retrieval IDs without proving they were the exact launch refs. A validly re-digested envelope replacement could therefore authorize a ref absent from the bound launch manifest. This is a same-slice S5 defect. The normative target is frozen and remains unchanged.

**Contract frozen before implementation:** `issueContextRefAuthorization` must receive the actual immutable launch `ContextManifest` used for the current `ExecutionBinding`. Under the current controller lock and owner/lease checks, it must reject unless the canonical manifest digest equals `ExecutionBinding.contextManifestDigest`, the durable envelope's verified `provenance.sha256` equals the manifest's `envelopeDigest`, and the full unique ref set in the manifest exactly equals the durable envelope retrieval set by `(refId, artifactPath, sourceDigest)`. Missing, duplicate, extra, omitted, or differently sourced refs fail closed. Grant entries are constructed only after these checks and current source bytes still match their source digests. The existing receipt continues to bind the current execution, context, and prompt digests; that receipt is not treated as proof of ref-set equality. The change does not alter S1 identity, S3 authority, retrieval capabilities, or the frozen target.

**Regression acceptance frozen before implementation:** (1) a self-consistent replacement envelope with an additional retrievable ref not in the launch manifest must be rejected; (2) a launch manifest whose ref ID is present but whose artifact path or source digest differs from the durable envelope must be rejected; (3) exact matching manifest/envelope/source evidence still issues only those refs. Run these tests first against the old implementation and record the expected red result before applying production changes; then require the same tests to pass. Tests must use deterministic fixtures and must not claim provider or adversarial certification.

### Remediation Tactical WorkGraph

| ID / owner | Mechanism | Dependencies | Bounded files / areas | Frozen invariant | Acceptance criteria |
| --- | --- | --- | --- | --- | --- |
| S5-R0 — LUNA_OWNED inspection and remediation contract freeze | HYBRID: semantic reading of the frozen S1/S3/S5 contracts with deterministic caller/source inspection | Accepted S3 checkpoint and declared S5 working tree; completed before source edits | This assessment; `docs/CORE_ARCHITECTURE_V2.md`; `src/workers/agentPrompt.ts`; `src/context/authorizationV2.ts`; `src/context/runtimeV2.ts`; existing authorization tests | Same-slice S5 only; preserve normative target and current S1/S3 authority semantics | Record actual callers, current binding checks, exact manifest/envelope comparison contract, and test criteria above before production source changes. |
| S5-R1 — FLASH_DELEGATABLE authorization mismatch regressions (tests only) | DETERMINISTIC | R0 contract frozen; separate Paseo worktree containing the declared S5 test/runtime snapshot | `tests/contextAuthorizationV2.test.ts` only | No production source, authority, schema, or documentation edits; no child agents/delegation; do not commit or push | Add fail-closed cases for an unbound extra ref, envelope-digest drift with unchanged refs, artifact-path mismatch, and source-digest mismatch, plus matching exact-evidence acceptance. Run the mismatch tests before production repair and report the old behavior's failing assertions. |
| S5-R2 — LUNA_OWNED review, integrate regression tests, and reproduce red | DETERMINISTIC | R1 worker output reviewed; test-only diff copied into the shared checkout | `tests/contextAuthorizationV2.test.ts` | Keep only tests within the frozen criteria; do not weaken existing validation | Confirm the regression tests fail specifically because the old issuer returns unauthorized refs or accepts a launch/envelope mismatch. |
| S5-R3 — LUNA_OWNED minimal issuance-contract repair | DETERMINISTIC | R0 frozen contract and R2 reproduced red; sequential with production implementation | `src/context/authorizationV2.ts`; `src/workers/agentPrompt.ts`; affected authorization test call sites only | Exact S1 launch manifest and S3 owner/read-lease gate remain authority; verify complete ref-set equality before grant construction; preserve source-byte and receipt validation | Pass R1 regressions and existing exact-ref/controller/session/retrieval/provenance tests. No compatibility path or model-issued authority. |
| S5-R4 — LUNA_OWNED integration review and evidence reconciliation | HYBRID: lead semantic acceptance plus deterministic reproducible gates | R3 complete | S5-affected status rows, Conformance, Ledger, assessment, and declared S5 files | Do not edit normative target, start another slice, commit, or push; do not promote deterministic fixtures to provider/adversarial certification | Run regression red then green; focused context/session/provenance tests; TypeScript, build, full non-system Vitest, and relevant deterministic system suites; record exact commands/counts and reconcile claims/evidence. |

### Delegation and ordering

R1 is a meaningful, safely bounded tests-only implementation unit assigned to one configured DeepSeek V4.1 Flash Implementer in the separate Paseo worktree `/home/james/.paseo/worktrees/0jx5pvzi/core-v2-s5-context-auth-manifest-regression` (workspace `wks_1e973a62648b706a`, agent `a3a1c3ee-9a21-4659-9bda-313e6c9985d0`). Immediately before delegation, Paseo `list_profiles` returned the named profile **DeepSeek V4.1 Flash Implementer** with provider `opencode`, model `opencode-go/deepseek-v4.1-flash`, mode `build`, thinking option `max`, and features `{ "auto_accept": true }`; those values were copied exactly to `create_agent`. The worker receives only R1 and must not create agents or delegate further (S5 rule). R0, R2, R3, and R4 remain LUNA-owned because they require contract control, source integration, and final evidence reconciliation. R1 → R2 → R3 → R4 are sequential; no parallel authority or implementation unit is opened.

The original S5 evidence above records the pre-remediation checkpoint; the current remediation result and reconciled evidence are recorded below.

**R1/R2 red-phase evidence (before production repair):** the reviewed Flash diff changes only `tests/contextAuthorizationV2.test.ts`. In the shared checkout, `npx vitest run tests/contextAuthorizationV2.test.ts --reporter=dot` exited 1 with `Test Files 1 failed (1)` and `Tests 4 failed | 4 passed (8)`. The failing cases all received `issued: true` from the old issuer: the self-consistent envelope authorized `evidence, smuggled`; the unchanged ref set with a rebuilt envelope digest authorized `evidence`; and separate launch `artifactPath` and `sourceDigest` mismatches each authorized `evidence`. The exact-match manifest/envelope/source case and the three existing lifecycle tests passed. This reproduces the bounded S5 defect before production changes.


### Remediation result and current evidence

The frozen contract was implemented in `src/context/authorizationV2.ts` and its launch caller in `src/workers/agentPrompt.ts`. The issuer receives the actual launch ContextManifest and, after the existing controller-owner, participant, and read-lease checks, verifies its canonical digest against the current ExecutionBinding, verifies `manifest.envelopeDigest` against the durable envelope's validated provenance digest, and verifies exact unique equality of the manifest and envelope ref sets by `(refId, artifactPath, sourceDigest)`. It rejects malformed, duplicate, missing, extra, or mismatched refs before grant construction. Existing safe-path and source-byte digest verification remains in force. S1 identity, S3 authority semantics, and the frozen normative target were not changed.

The configured Flash worker changed tests only in the separate Paseo worktree recorded above. Lead review integrated the regression cases into the shared checkout and confirmed they failed against the old issuer before the production repair. After repair, all exact-match and mismatch cases pass.

**Exact current validation evidence (2026-09-24):**

- Red first: `npx vitest run tests/contextAuthorizationV2.test.ts --reporter=dot` — before production repair, exit 1; 1 file failed, 4 tests failed and 4 passed (8 total). The four failures were the added envelope ref, changed envelope digest with unchanged ref tuples, artifact-path mismatch, and source-digest mismatch.
- Regression green: `npx vitest run tests/contextAuthorizationV2.test.ts --reporter=dot` — exit 0; 1 file passed, 8 tests passed.
- Focused S5 context/session/provenance command: `npx vitest run tests/contextArchitectureClose.test.ts tests/contextAuthorizationV2.test.ts tests/contextEfficiency.test.ts tests/contextGuard.test.ts tests/contextRuntimeV2.test.ts tests/agentPromptStructuredResultProvenance.test.ts tests/structuredResultGateway.test.ts tests/paseo/sessionBinding.test.ts tests/paseoPendingExecutionIdentity.test.ts tests/paseoInitialTurnBarrier.test.ts tests/distributedExecutionIdentity.test.ts tests/reviewerSerializationRecovery.test.ts --reporter=dot` — exit 0; 12 files passed, 125 tests passed.
- TypeScript: `npx tsc -p tsconfig.json --noEmit` — exit 0.
- Build: `npm run build` — exit 0; built `release-1790293375971-2416336-1fc6226d`.
- Full non-system Vitest: `npx vitest run --exclude 'tests/system/**' --reporter=dot` — exit 0; 152 files passed, 1 skipped; 835 tests passed, 7 skipped.
- Relevant deterministic system lane: `npx vitest run tests/system/aehAdversarialE2E.test.ts tests/system/aehConcurrencyCampaign.test.ts tests/system/aehScenarioMatrix.test.ts --reporter=dot` — exit 0; 3 files passed, 45 tests passed. These are deterministic fixtures; no adversarial certification is claimed.

Evidence lanes are CONTRACT, UNIT, INTEGRATION, BUILD, and SYSTEM_DETERMINISTIC. No REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or PACKED_E2E certification is claimed. Status rows 24–27, Conformance's Context Runtime v2 claim, and Ledger finding AEH-V2-0011 now reflect the current regression and gate results. `docs/CORE_ARCHITECTURE_V2.md` remains unchanged. No slice beyond S5 was started; no commit or push was made.

### Campaign Director independent gate — SLICE_ACCEPTED

The Campaign Director reviewed the remediation report, complete declared S5 diff, the authorization issuer and launch caller, the regression tests, all affected status/conformance/ledger entries, and the S5 gate. The immutable launch-manifest digest, verified envelope provenance digest, and exact unique `(refId, artifactPath, sourceDigest)` set are checked before grant construction; source-byte verification remains in place. The negative tests exercise self-consistent extra refs, envelope digest drift, and path/source-digest substitution. No normative target change or unresolved integration conflict was found. `git diff --check` passed, and `docs/CORE_ARCHITECTURE_V2.md` is unchanged.

Independent reruns on 2026-09-24: focused context/session/provenance — 12 files / 125 tests passed; `npx tsc -p tsconfig.json --noEmit` — exit 0; `npm run build` — exit 0 (`release-1790294205611-17-b4fe25cb`); full non-system Vitest — 152 files passed, 1 skipped; 835 tests passed, 7 skipped; relevant deterministic system suites — 3 files / 45 tests passed. The full non-system suite initially encountered sandbox loopback/TSX IPC `EPERM`; rerunning the same authorized gate with local process/loopback access passed. S5 is accepted on its deterministic gate only. Provider, browser, live OPA, adversarial, and packed end-to-end certification remain pending. The next dependency-ready slice is S4. The Director will checkpoint this accepted S5 working tree; this record is written before that commit.
