Document Notarius cache reuse workflow

This commit is contained in:
2026-09-16 18:14:40 +00:00
parent 8ff74e1251
commit a0312ed034
11 changed files with 228 additions and 31 deletions

View File

@@ -65,6 +65,9 @@ Notarius is composed only when extraction is enabled; the extract stage owns
prepared reference resolution, receipt, bundle, and configured-lane policy.
The adapter validates the ordered selector/absolute-path pairs and is the sole
owner of serializing them as repeated `--reference` arguments before `--json`.
It also serializes an effective resume request as `--resume` and a nonempty
validated chunk mode as `--chunk_cache <mode>`. The extract stage, not the
adapter, resolves force precedence before constructing the request.
Its request may also carry one already-resolved opaque ConfigKit secret
reference. Only the Notarius adapter reads that value, rejects an unresolved or
empty programmatic reference, and maps the exact configured name/value into the

View File

@@ -19,13 +19,15 @@ procedures belong in [Operations](../operations.md).
1. resolves the final trimmed transcript from the shared artifact catalog;
2. resolves every configured prepared reference through the shared
manifest-authoritative identity resolver before creating run-local output;
3. streams each verified reference into an invocation-local snapshot and
rejects any source change observed while copying;
3. resolves each reference to a stable content-addressed session snapshot,
reusing verified bytes or reconstructing a missing/corrupt regular entry,
and rejects any source change observed while copying;
4. fingerprints the byte- and provenance-bearing Notarius invocation evidence,
including sorted reference identities;
5. creates a run-local staging directory and invokes the injected
`notarius.Runner`, carrying the optional already-resolved credential
reference without exposing its value;
reference without exposing its value, plus effective resume and chunk-cache
controls after force precedence;
6. revalidates the reference snapshots, then validates the v2 successful
receipt, confined index, management documents, configured required lane
descriptors, validation summaries, and regular payload files;
@@ -42,14 +44,22 @@ Narratio run ID, the resolved trimmed-input identity, and invocation
fingerprint. The input identity binds the exact transcript bytes, canonical
source ID, producer stage/output/run identity, and resolution provenance.
Reference metadata contains only selector, source ID, canonical session-relative
path, checksum, and size; adapter requests receive selector and absolute
invocation-local snapshot path, never payload contents. Snapshot bytes must
path, checksum, and size; adapter requests receive selector and absolute stable
snapshot path, never payload contents. Snapshot bytes must
match the prepared identity both before and after Notarius runs, so a concurrent
prepared-file replacement cannot make recorded provenance describe different
bytes from those supplied to Notarius.
Validation completes before the single durable rename, so a rejected result
cannot expose a partial durable bundle.
Snapshot destinations are owned by `internal/artifacts` under
`cache/notarius/references/<sha256>/<prepared-basename>`. They contain no run ID
or timestamp and can be reused across invocations. Identical bytes and basename
share a verified file; changed bytes select a different digest directory.
Unsafe path components and non-regular conflicts fail without recursive
deletion. Snapshot paths are delivery details and never become manifest output
evidence or prerequisites for reuse of an already complete extract bundle.
Any executed extraction outcome that replaces a different effective outcome
marks succeeded analysis and delivery dependents stale. Render is an independent
sibling and remains current. Repeating the same disabled self-skip with no
@@ -74,7 +84,10 @@ sorted prepared-reference identities, pipeline identity, and configured output
contracts. The same reference helper and transcript identity are resolved again
for artifact evidence, so changing current transcript bytes, reference bytes,
or producer identity makes the prior extraction obsolete. Operational runner
settings do not invalidate otherwise current durable evidence.
settings do not invalidate otherwise current durable evidence. This includes
`resume` and `chunk_cache`; changing them alone cannot make a complete
extraction non-resumable. If extraction executes, ordinary runs forward the
configured values, while force disables resume and selects `refresh`.
A valid prepared-reference change makes extraction non-resumable. Missing,
unsafe, or checksum-inconsistent prepared evidence is a hard validation error
@@ -115,9 +128,13 @@ recovery.
`internal/stage/extract_resume.go`,
`internal/stage/extract_test.go`,
`internal/stage/semantic_contracts_delivery.go`
- Stable snapshot materialization and verification:
`internal/stage/extract_references.go`
- Subprocess boundary: `internal/adapters/notarius/subprocess.go`,
`internal/adapters/notarius/subprocess_test.go`
- Catalog hydration: `internal/artifacts/extraction_catalog.go`,
`internal/artifacts/extraction_catalog_test.go`
- Composition and downstream behavior: `internal/app/runner_test.go`,
`internal/stage/analyze_test.go`, `internal/stage/publish_test.go`
- Opt-in Notarius v0.7.0 production-boundary coverage:
`internal/stage/notarius_cache_integration_test.go`

View File

@@ -37,6 +37,12 @@ validates and renames the external bundle to the unique immutable Notarius
bundle path supplied by `internal/artifacts`. Exact physical paths belong in
[Operations](../operations.md#extraction-workflow).
Prepared references delivered to Notarius use session-durable paths owned by
`internal/artifacts` beneath
`cache/notarius/references/<sha256>/<prepared-basename>`. These verified
content-addressed snapshots are outside run-local layout, survive ordinary
published-run cleanup, and are excluded from publish and restore enumeration.
## Locking
`artifacts.LocalStore` enforces the single-writer session lock via an