Plan shared stage checkpoint mechanics

This commit is contained in:
2026-09-22 14:54:34 +00:00
parent 792cbd0677
commit 9f2cfddce6
2 changed files with 230 additions and 372 deletions

View File

@@ -1,416 +1,272 @@
# Implementation Plan: Speaker Checkpoints and Replayable Uploads # Implementation Plan: Shared Stage Checkpoint Mechanics
Status: Implementation complete; Stages 1-7 complete. Status: Not started.
## Scope and Instructions ## Scope and Authority
Implement the agreed follow-up to v2.2.0: consolidate the thin checkpoint
mechanism shared by transcribe and analyze without combining their progress
models or changing operator behavior. There is no separate consolidation
roadmap in this checkout. This plan uses the accompanying discussion as its
scope: persistence is already centralized; candidate isolation and
save-before-promotion are the duplication to remove. The completed
[transcription roadmap](transcribe-checkpoints-and-replay.md) provides background,
not authorization to reimplement transcription or HTTP replay. The preceding
completed implementation plan remains in Git history.
Implement [the feature roadmap](transcribe-checkpoints-and-replay.md). This plan
replaces the completed Notarius-profile plan; that work remains in Git history.
Complete the numbered stages in order, one bounded implementation prompt per Complete the numbered stages in order, one bounded implementation prompt per
stage for gpt-5.6-terra. Record completion and actual verification results here. stage for gpt-5.6-terra. Record completion and actual verification results here.
Intermediate stages are not a finished feature and must not be released alone. This planning pass changes documentation only; the steps below are for the
subsequent implementation sprint.
Before working, read `AGENTS.md`, `docs/development.md`, all three policies in Before each stage, read `AGENTS.md`, `docs/development.md`, and all three
`docs/policy/`, the roadmap, and the stage-specific preparation below. Use the documents in `docs/policy/`: architecture, documentation, and testing. Use
maintained Go 1.26 toolchain. Preserve operator choices, speaker filenames, maintained Go 1.26. Prefer graph tools for code discovery; inspect current
bounded concurrency, streaming, and existing credentials/settings. Do not add files if their index is stale. Read the stage-specific preparation below,
new commands, dependencies, service requirements, arbitrary speaker limits, rather than unrelated historical roadmaps.
strict transcript schemas, or a generic checkpoint framework. No release or
upstream service work is part of this plan.
## Settled Design Decisions ## Target and Scope Boundaries
### Progress Representation and Recovery The existing `internal/manifest.Store` and `LocalStore` remain the session
manifest persistence interface and implementation. Encoding, schemas,
validation, normalization, and atomic file replacement remain unchanged.
There is no new store, journal, file format, dependency, command, configuration,
or migration.
Add optional `transcribe_state_version` (version 1) and `transcribe_speakers` One private helper in `internal/stage` owns candidate preparation, calling the
to `manifest.StageRecord`. The latter is a map keyed by the existing derived existing store, and promoting the selected stage record only after successful
speaker ID, containing completed results only. Absence means that work must be save. The existing typed analyze/transcribe checkpoint methods become thin
done; a separate pending/running/failed state machine is unnecessary. Aggregate callers. This is execution plumbing: keep it beside the callers rather than
stage status/error remains authoritative for invocation failure. A versioned moving orchestration into the storage package or exporting a reusable API.
empty collection is meaningful: no earlier completions are reusable.
Use a typed `TranscribeSpeakerRecord` with fingerprint version/digest, Keep these responsibilities stage-owned:
session-relative audio path, audio SHA-256 and size, one `ArtifactRecord` output
of kind `transcript_raw`, output size, and producer run ID. The output carries
its session-relative canonical path and SHA-256. No new transcript source IDs,
synthetic schema contracts, timestamps, or historical attempt collections are
needed. Existing aggregate diagnostics can retain attempt/status/duration data
for work performed in the current invocation; do not fabricate that information
for reused outputs.
Keep the session-manifest schema identifier unchanged for this optional additive - Transcribe's completed-only speaker map, evidence checks, bounded concurrency,
state; validate supported per-stage versions and retain strict decoding. New mutex, checkpoint-failure latch, cancellation, pre-dispatch force reset,
Narratio reads old manifests. Old binaries may reject the added fields under output installation, and aggregate repair.
their existing strict decoding; document this downgrade limitation. - Analyze's dispositions, selection, dependencies, catalog updates, working and
committed collections, failure reconciliation, invalidation, and outputs.
- Each stage's decision whether/when to checkpoint, typed validation, contextual
errors, and post-save working-state updates.
- Runner aggregate transitions, session locking, terminal failure saves, and
fixed dependent-stage invalidation.
For pre-feature manifests with no speaker evidence, regenerate transcription Do not introduce a common subtask model, scheduler, transaction manager, global
once when transcribe is next selected, even if the aggregate status succeeded. mutex, automatic retry, additional checkpoint, or batching policy. Do not
Old records do not bind individual transcript results to the audio bytes they migrate publish/restore, planning clones, cleanup, or ordinary runner saves:
used. Do not guess that association from current files, old `per_file` metadata, their operations are not the demonstrated duplication. Leave HTTP replay and
or abandoned run directories. This is the conservative upgrade path, not a new all adapters unchanged.
general migration system. No other stage or unselected range is forcibly run.
### Reuse and Persistence ## Selected Design
Compute a versioned per-speaker fingerprint from speaker ID, session-relative ### Shared Save-and-Promote Operation
audio identity, actual audio SHA-256/size, and the existing transcribe semantic
configuration fingerprint (URL and language). Do not include invocation ID,
absolute workspace root, concurrency, retry settings, timeout, or credentials.
Stream input hashing with cancellation; do not trust stored input checksums
alone. Recheck input bytes before accepting a new result to catch changes
observed during transcription; fail that job without checkpointing mismatched
evidence. This is not a hostile-filesystem snapshot or concurrent-editor
transaction guarantee and requires no extra audio copies or file locks.
Validate reusable output identity, canonical location, JSON, checksum, and size. Add `internal/stage/checkpoint.go` with a private helper shaped as:
Use existing confinement/file primitives. Missing or corrupt regular outputs
become work; unsafe paths/non-regular conflicts follow existing safety errors,
not recursive deletion. Missing required audio is an input error, not permission
to reuse an unverifiable transcript. Removed inputs are removed from the current
collection and aggregate outputs without deleting arbitrary files.
One stage-owned mutex serializes candidate collection construction, manifest ```go
save, and installation of committed state. Save a detached candidate and only func checkpointStageCollection(
then update live authority, following analyze's pattern. Workers must not mutate ctx context.Context,
the live manifest outside that serialized section. Keep expensive HTTP work store manifest.Store,
outside the mutex. Stop dispatch/acceptance on checkpoint failure, join workers path string,
before return, and preserve the last successfully committed collection. live *manifest.Manifest,
stageName string,
update func(*manifest.StageRecord) error,
) error
```
Before launching requests, durably checkpoint the reconciled collection. For Names are implementation choices, not public contracts. Use a function and
force, that collection is empty: this prevents later ordinary retries from small copying helpers, not a factory/options/interface framework. The two
resurrecting pre-force outputs. If this initial save fails, do not start requests existing typed checkpoint methods are the only production callers. They pass
or overwrite outputs. Force again clears the current collection again; ordinary the already-resolved manifest path; the helper does not discover paths or
retry retains replacements already checkpointed after the previous force. inspect configuration.
Save each validated result at its canonical raw-transcript location before The helper performs this sequence:
checkpointing it. Run-local files remain staging/diagnostics, not reuse
dependencies. Bytes left by an unsuccessful checkpoint have no completion
authority. A worker may checkpoint a completed result using the parent stage
context even if a sibling canceled the work context; parent cancellation or a
checkpoint failure may prevent that save. Guarantee retention of committed
work, not all responses that happened to finish during cancellation.
### Multipart Replay Mechanism 1. Check required live/store inputs without mutation. Copy a checkpoint
candidate and select a copied target record. If absent, provide an empty
candidate record to the caller, without inferring its name, status, or time.
2. Apply the supplied update to that candidate record only. On update error,
return without saving or promoting. The typed caller validates and fills
its version, collection, and stage-owned initialization/timestamps.
3. Invoke `Store.Save` once with the supplied context. Preserve error identity
through `%w` and useful path context. On error, do not promote, retry,
reload, or synthesize aggregate failure.
4. After success, install the saved target record and candidate manifest
`UpdatedAt` into live state. Preserve an existing target record's pointer
identity by assigning through it. Allocate missing live maps/records only
after success. Do not replace the entire live manifest or overwrite other
stage records. Copy mutable checkpoint collections on promotion so saved
candidates and caller working collections cannot mutate live progress.
Replace the pipe/producer upload mechanism with a small pull-based multipart Callers serialize the whole operation. Transcribe's current mutex does so;
body backed by the audio reader. Use `mime/multipart.Writer` to encode a bounded analyze executes artifacts sequentially. Do not add a lock to the helper or
metadata prefix and suffix around the file payload: capture the prefix after promise support for concurrent independent manifest writers. Atomic replacement
`CreateFormFile`, then encode the language field and closing boundary as the visibility does not become power-loss durability or a transaction covering
suffix. Generate this framing once per application attempt. Each body combines transcript files.
fresh readers for prefix, freshly opened audio, and suffix using `io.MultiReader`.
No full-audio buffer or temporary multipart spool is needed.
The initial body and `Request.GetBody` use the same factory/framing. Each owns ### Bounded Candidate Isolation
an independent audio handle. Give bodies idempotent Close and context-triggered
closure (for example `context.AfterFunc`), without holding a mutex across a
blocking Read. Read/Close coordination must release a blocked source and retain
genuine source-read/close errors while not treating deliberate early closure as
source failure. The attempt owns cleanup of all created bodies, including
replays; reject new body creation after its context expires.
Remove the old producer `Wait` dependency. Track body-local source errors with Factor `cloneManifestForTranscribeCheckpoint` into a shared checkpoint-scoped
synchronization so a swallowed upload read error cannot become a successful copy helper in the new file. Use ordinary typed copies and existing
transcript, and a deliberately closed original body cannot poison a successful `manifest.CloneAnalyzeArtifactCollection` and
replay. Keep the outer retry policy and per-attempt context. Do not add replay `manifest.CloneTranscribeSpeakerCollection`. Do not introduce reflection
counters to the user retry budget or depend on Go's exact internal retry count. cloning, JSON round trips, or a public general-purpose clone API.
Preserve ordinary HTTP redirect policy; body replay can also enable Go's
body-preserving redirects. Do not introduce a new redirect restriction.
## Stage 1: Add Durable Speaker Completion State Copy the manifest value, stages map, stage records, and mutable collections
touched by these updates or `LocalStore.Save`. In particular, Save normalizes
analyze collections throughout the candidate, including on a transcribe
checkpoint. Clone those collections for every stage so normalization cannot
leak into live state on failure. Preserve the existing bounded copies of
input/artifact and output/log/config slices. Retain nil entries for existing
validation to reject, rather than silently repairing invalid manifests.
Status: Complete. This is not arbitrary deep isolation of metadata. Neither the two update
callbacks nor the current Store mutates `map[string]any` metadata, so it may
remain shared read-only. Document that boundary and inspect every current
normalization write. Use existing typed clones for nested checkpoint evidence;
do not restrict legal metadata values or alter their Go/JSON representations
to simplify copying.
Completion marker (2026-09-21): Added optional versioned transcribe-owned ### Differences to Preserve
per-speaker completion records, strict load/save validation, isolated collection
cloning, legacy aggregate-only compatibility, and lifecycle-preservation
coverage. Verified with `go test ./internal/manifest` and `git diff --check`.
Preparation: `internal/manifest/manifest.go`, `store.go`, `analyze_state.go`, - Analyze still requires an existing stage record. Its callback changes only
their tests, and `docs/internal/manifest.md`. analyze version/collection, not transcribe-style initialization or timestamp
policy. It refreshes working and committed collections only after success.
- Transcribe retains missing-record initialization for direct stage calls:
its typed caller supplies name, running status, and initial timestamps when
needed. Its callback retains the existing stage `UpdatedAt` update. Copying
an existing record must not clear aggregate fields or invoke lifecycle
transitions.
- Existing transcribe live-record initialization is incomplete on the unusual
missing-record path; promoting the actual successfully saved candidate makes
its identity agree with disk. This does not add a new missing-record
rejection or change ordinary runner status transitions.
- Analyze rollback/failure handling and transcribe failure latching remain in
their execution owners. The shared operation never invents failed items.
- Preserve starting-checkpoint elision, force/reset semantics, empty versioned
collections, evidence and producer identity, legacy transcription handling,
supported zero-sized audio, and aggregate repair without HTTP calls.
- Do not freeze exact error wording, wall-clock values, copy choreography, or
iteration order into new contracts. No new schema, limits, filename rules,
statuses, or compatibility guarantees are authorized.
1. Implement the additive fields, record type, version constants, validation, ## Stage 1: Add the Shared Checkpoint Operation
and deep-copy helper described above in a focused transcribe state file.
2. Validate ownership by the transcribe stage, supported versions, required
completed-record evidence, output kind, safe relative paths, SHA-256 values,
sizes, and producer identity consistency. Use existing speaker IDs as-is;
do not apply analysis-artifact key regexes or copy unrelated analysis limits.
3. Integrate validation with both load and save. Preserve speaker state through
running/failed/stale transitions and aggregate result clearing. Generic stage
invalidation does not discard checkpoints; reconciliation checks their data.
4. Test round-trip populated and versioned-empty collections, old manifests
without fields, malformed evidence, deep-copy isolation, and lifecycle
preservation. Avoid imposing JSON formatting or map order as contracts.
Verify: `go test ./internal/manifest`. Status: Pending.
Done when durable storage can represent partial completion independently of Preparation: `internal/manifest/store.go`, `manifest.go`, `analyze_state.go`,
aggregate status, without activating new stage behavior yet. `transcribe_state.go`; `internal/stage/analyze_execution_state.go`,
`transcribe_execution_state.go`; `docs/internal/manifest.md`.
## Stage 2: Implement Read-Only Speaker Reconciliation 1. Add the private helper and bounded candidate copying described above.
Leave existing production checkpoint paths unchanged until their migration
stages. Do not duplicate manifest record-validation rules.
2. Add focused `internal/stage/checkpoint_test.go` coverage with real
`manifest.LocalStore` and temporary directories for successful save/reload.
Use a small Store wrapper only to observe/fail the persistence boundary.
3. Prove live progress remains old during Save, matches saved progress after
success, and does not change on update, validation, or save failure. Include
both typed collections and a transcribe checkpoint alongside analyze
records subject to save-time normalization. A failure test must exercise
candidate timestamp/normalization mutation before returning an error, not
merely fail before any work.
4. Prove unrelated stage/aggregate data survives, existing target references
observe updates, missing-target creation is invisible on failure, and later
caller collection edits do not mutate committed records. Preserve context
and error identity. Test the supported copy/write boundary, not arbitrary
malicious Store mutation.
Status: Complete. Verify: `go test ./internal/manifest ./internal/stage` and
Completion marker (2026-09-21): Extracted shared deterministic transcribe job
construction and added read-only, cancellation-aware per-speaker input and
output reconciliation. Covered unchanged reuse, semantic and operational
settings, changed/added/removed inputs, legacy and forced work, missing/corrupt
outputs, unsafe evidence, and missing audio without connecting runtime skip or
checkpoint behavior. Verified with
`go test ./internal/stage ./internal/manifest` and `git diff --check`.
Preparation: `internal/stage/transcribe.go`, `semantic_contracts_initial.go`,
`analyze_reconciliation.go`, `extract_resume.go`, their tests, and the artifact
path/integrity helpers they use.
1. Extract the existing audio discovery/job construction into a shared,
transcribe-local reconciliation path. Preserve current input selection and
speaker identity behavior. Build sorted jobs and classify completed results
against the current input set and evidence defined above.
2. Add pure fingerprint construction and streamed input/output inspection.
Resolve manifest input paths with existing session-relative/absolute reading
helpers; do not change input precedence or assume the current working directory
is the session root. Newly written checkpoint paths are always relative.
3. Return current records and work-to-do without writing files or manifest state.
Force selects all work and an empty starting collection; old unversioned
transcribe state supplies no reusable records. Do not connect aggregate skip
validation until execution can produce the new evidence in Stage 3.
4. Test unchanged reuse; one changed audio among unchanged speakers; URL/language
changes; operational-setting exclusions; added/removed speakers; missing and
corrupt regular outputs; unsafe record paths; missing audio; and force.
Prefer compact tables and real temporary files over mocked hashing.
Verify: `go test ./internal/stage ./internal/manifest`.
Done when one read-only decision path describes exactly which speakers need
work, without altering existing runtime behavior prematurely.
## Stage 3: Execute and Checkpoint Speakers Incrementally
Status: Complete.
Completion marker (2026-09-21): Connected reconciliation to bounded execution,
durably checkpointed the starting and per-speaker collections through detached
manifest candidates, preserved committed work across failures and cancellation,
rebuilt deterministic aggregate outputs from reusable evidence, and added
speaker-aware resume validation. Updated production-style fixtures and covered
serialized concurrent saves, force/retry replacement, input rechecks, save
failures, partial recovery, and aggregate reconstruction. Verified with
`go test ./internal/stage ./internal/app`, `go test -race ./internal/stage`, and
`git diff --check`.
Preparation: `internal/stage/transcribe.go`, `analyze_execution_state.go`,
`run_local.go`,
`internal/app/runner.go`, and transcribe/checkpoint test fixtures.
1. Wire reconciliation into Run, require the existing ManifestStore collaborator,
and checkpoint the starting collection before launching HTTP work. Implement
detached candidate saves and serialized live-state updates. Audit Store.Save's
normalization side effects so a failed candidate save cannot mutate shared
authority; use detached records/collections wherever it may write.
2. Retain bounded errgroup execution and fail-fast cancellation. Check the work
context before beginning each queued request. After success, validate adapter
output path/JSON, recheck audio evidence, materialize canonical bytes, compute
output evidence, and checkpoint the new record. Check a checkpoint-failure
latch under the serialization lock before accepting further completions.
3. Preserve committed successes on all error/cancellation exits, including
starting-save, canonical-write, per-speaker-save, and aggregate-save failures.
Return meaningful speaker/error context and let the runner retain its existing
terminal failure behavior. Do not report partial aggregate outputs as success.
4. Assemble final outputs in deterministic speaker order from the reconciled
completed records, including reused speakers, with original producer IDs and
checksums. If all records are current but aggregate status is failed/running,
finish without WhisperX calls and rebuild the aggregate success.
5. Implement `transcribeStage.ValidateResume` using the same read-only
reconciliation, so a succeeded stage cannot skip changed audio or invalid
transcripts. Retain the existing URL/language semantic check and runner
invalidation behavior; no whole-config digest invalidation.
6. Adapt existing tests that require no canonical partial files: the new contract
permits checkpointed files but forbids aggregate success. Update fixtures to
provide real manifest stores and records rather than a production fallback
that silently disables checkpointing.
7. Test A/B completion followed by C failure and D unattempted (concurrency 1),
reload and retry only C/D; concurrently finishing speakers without lost saves;
failure and cancellation after a checkpoint; save failure retaining earlier
authority; and force failure followed by ordinary retry without old-result reuse.
Verify: `go test ./internal/stage ./internal/app` and
`go test -race ./internal/stage`. `go test -race ./internal/stage`.
Done when incremental progress and aggregate success work together safely under Done when the shared integrity boundary is tested without changing existing
ordinary, concurrent, canceled, and forced execution. stage behavior. Temporary copying overlap is removed in Stage 3.
## Stage 4: Verify Runner, Merge, and Durable Lifecycle Integration ## Stage 2: Migrate Analyze
Status: Complete. Status: Pending.
Completion marker (2026-09-22): Added real-persistence application coverage for Preparation: `internal/stage/analyze_execution_state.go`,
cross-invocation retry, interrupted running state, failed aggregate saves, `analyze_checkpoint_test.go`, `analyze_failure_test.go`,
force/reset persistence, downstream invalidation, and one-time legacy upgrade. `internal/app/analyze_checkpoint_recovery_test.go`, and
Versioned merge input discovery now requires aggregate success and an exact `docs/internal/stage-analyze.md`.
match between current speaker evidence and declared aggregate outputs, while
legacy/manual directory fallback remains available. Restore coverage preserves
session-relative speaker paths, leaves unpublished raw transcripts absent, and
canonical transcripts remain reusable after run-directory cleanup. Verified
with `go test ./internal/app ./internal/stage ./internal/manifest`,
`go test -race ./internal/app`, and `git diff --check`.
Preparation: `internal/app/runner.go`, `runner_runtime.go`, 1. Replace manual candidate copying, saving, and live promotion in analyze
`analyze_checkpoint_recovery_test.go`, `runner_terminal_failure_test.go`, with the shared helper. Retain typed validation, required-record checking,
`restore_manifest_paths.go`, `internal/stage/merge.go`, and existing publish, the typed update callback, and committed/working collection refresh.
restore, and cleanup tests routed from `docs/internal/overview.md`. 2. Preserve starting-checkpoint elision, selected-artifact behavior, catalog
sequencing, dependent invalidation on changed output, and rollback after
checkpoint failure. Do not change the runner's terminal failure save or
joined errors.
3. Run existing tests before adding coverage. Retain tests proving earlier
artifacts survive failure/cancellation, unsaved output bytes gain no
authority, selected/forced work preserves appropriate unrelated artifacts,
and interrupted runs recover across store reloads. Add only missing
integration coverage; do not duplicate Stage 1's isolation matrix.
1. Add application-level tests with real manifest persistence and fake WhisperX: Verify: `go test ./internal/stage ./internal/app ./internal/manifest` and
failed stage across separate invocations; all speakers committed before a `go test -race ./internal/stage ./internal/app`.
failed aggregate save; interrupted running state; force/reset persistence;
and unchanged upstream/succeeded downstream invalidation behavior. Inject
errors at durable boundaries, not every private method call.
2. Close `discoverRawTranscripts`' directory-scan fallback for versioned
transcribe state: incomplete aggregate state is not usable, and completed
state consumes only declared current aggregate outputs. Keep legacy/manual
fallback behavior for manifests without the new state rather than banning
previously allowed standalone workflows. Test that partial files and removed
speaker files cannot leak into a versioned merge, including direct stage calls.
3. Audit manifest copies/projections, remote manifest publication, restored path
handling, and cleanup. Preserve relative speaker paths on restore and verify
they resolve under the destination session, not the producer machine. Reuse
existing confinement validation, not a second path-policy layer.
4. Add focused round-trip coverage through existing publish/restore fixtures for
speaker records. Restored outputs absent from the published selection are
non-reusable locally; do not silently expand remote publication to every raw
speaker or upload audio as a new default. Removing run directories must not
remove canonical checkpointed transcripts or make reuse require run files.
5. Verify old succeeded manifests are readable and rerun once when transcribe is
selected, then reuse normally. Keep old manifests usable for other commands
and bounded ranges. Do not automatically import old unrecorded JSON files.
Verify: `go test ./internal/app ./internal/stage ./internal/manifest` and Done when analyze delegates persistence choreography while progress, artifact
`go test -race ./internal/app`. selection, and failure behavior remain unchanged.
Done when the new authority survives runner and storage lifecycles and cannot ## Stage 3: Migrate Transcribe and Remove Duplicate Machinery
expose incomplete transcripts downstream.
## Stage 5: Implement Replayable Streaming Multipart Bodies Status: Pending.
Status: Complete. Preparation: `internal/stage/transcribe_execution_state.go`, `transcribe.go`,
`transcribe_resume.go`, `transcribe_checkpoint_test.go`,
`transcribe_reconciliation_test.go`,
`internal/app/transcribe_checkpoint_recovery_test.go`, and
`docs/internal/stage-transcribe.md`.
Completion marker (2026-09-22): Replaced the producer-goroutine/pipe upload 1. Delegate `checkpointCollectionLocked` to the shared helper, retaining
with bounded standard-library multipart framing and independently opened, typed initialization/update and successful working-collection refresh.
pull-based streaming bodies shared by the initial request and `GetBody`. Keep the existing mutex across candidate creation, save, and promotion.
Bodies are tracked, idempotently closed, canceled with their attempt context, Do not move HTTP/hashing work into that critical section or add a lock.
and retain genuine read/close failures without treating an intentionally 2. Preserve failure-latch checks and worker joining, the parent versus worker
abandoned original as a replay failure. Added focused coverage for identical context distinction, force's durable empty starting collection, and
framing, partial-read restart, simultaneous handles, source failures, and ordinary retry's reuse of committed forced replacements.
blocked-reader release. Verified with `go test ./internal/adapters/whisperx`, 3. Remove `cloneManifestForTranscribeCheckpoint`. Its current use by a
`go test -race ./internal/adapters/whisperx` (including ten repeated runs), and read-only resume test is not a reason to retain duplicate production code.
`git diff --check`. Use an independent test snapshot or the shared bounded copy where sufficient,
without exporting a helper for tests.
4. Retain existing real-store and concurrency coverage: no lost completions;
pre-dispatch save failure prevents requests/canonical writes; later failure
retains prior authority; interrupted and aggregate-save recovery; failed
force followed by retry; run-directory independence; zero-sized audio;
removed/mismatched aggregate repair without additional WhisperX requests
or removed-speaker merge input. Do not weaken these regression assertions.
Preparation: `internal/adapters/whisperx/http.go`, `http_test.go`, Verify: `go test ./internal/stage ./internal/app ./internal/manifest` and
`docs/integrations/whisperx.md`, and the installed Go net/http Request/GetBody `go test -race ./internal/stage ./internal/app`.
contract. Keep this change inside the WhisperX adapter.
1. Implement the pull-based framing/body factory described above, using standard Done when both stages use one save/promote implementation and one bounded
multipart encoding rather than hand-escaped MIME headers. Add no new payload copy implementation, with their different models and synchronization intact.
cap or upload-size setting. Preserve file-before-language wire semantics and
the existing form names.
2. Install the initial body and GetBody with the same per-attempt boundary.
Track and close all opened bodies even if a custom transport returns early;
make cleanup/context callbacks safe against overlapping reads and closes.
A GetBody failure must be returned, not converted into an empty upload.
3. Replace pipe-specific Wait/error paths in doTranscribeAttempt with source-error
handling appropriate to pull bodies. Genuine source errors remain visible;
intentional closure during transport retry is not itself a source error.
Keep response bounds, JSON validation, output installation, retry
classification/delay, cancellation, and application attempt reporting intact.
4. Remove obsolete producer-only helpers rather than retaining two upload paths.
Adapt tests tied to producer goroutines to assert streaming and resource
release, which are the actual contracts.
5. Add focused body tests: original/replayed bytes parse identically; replay
after partial consumption starts from the beginning; simultaneous bodies use
independent handles; source failures propagate; Close/cancel releases a
blocked reader without deadlock. Use synthetic audio and gated readers.
Verify: `go test ./internal/adapters/whisperx` and ## Stage 4: Document, Review Scope, and Validate
`go test -race ./internal/adapters/whisperx`.
Done when replay is available without whole-file buffering or producer-lifetime Status: Pending.
ambiguity, and all existing HTTP behavior tests still pass.
## Stage 6: Verify Replay Through the HTTP Adapter Boundary Preparation: `docs/internal/manifest.md`, `docs/internal/stage-transcribe.md`,
`docs/internal/stage-analyze.md`, documentation/testing policies, and the
implementation diff against this plan's starting commit.
Status: Complete. 1. Update manifest internals to identify the shared checkpoint owner and copy/
save/promote boundary. Stage guides retain stage-specific policy and link
Completion marker (2026-09-22): Added a deterministic transport regression to that explanation instead of duplicating shared implementation details.
that abandons a partially consumed original request, obtains and parses a fresh Do not rewrite immutable release notes or announce new operator behavior.
multipart replay from `GetBody`, and installs the returned transcript within one 2. Assess tests collectively. Consolidate genuinely duplicated persistence
application attempt. Covered replay open/read failures, early responses, mechanics coverage, but retain stage/application recovery tests proving
parent cancellation during replay, and disposal of every opened source, while correct integration. Avoid test-count/LOC quotas. Record a brief summary
proving an abandoned original does not poison a valid replay. Added a real HTTP here of duplicated mechanisms removed and thin wrappers retained.
retry-exhaustion integration that durably retains an earlier speaker and resumes 3. Confirm no schema, validation policy, scheduling, force, concurrency, output
only the uncommitted speakers. This controlled boundary verifies Narratio's selection, stage order, adapter call, or configuration change slipped into
replay contract, not recovery from every real proxy protocol failure. Verified the refactor. Additional restrictions require a separate decision, not
with `go test ./internal/adapters/whisperx ./internal/app ./internal/stage`, tests that retrospectively declare new policy.
`go test -race ./internal/adapters/whisperx ./internal/app ./internal/stage`, 4. Run all gates below and record actual results. Mark this plan complete only
ten repeated race-enabled adapter runs, and `git diff --check`. when both migrations and verification pass. Do not release, tag, push, delete
other roadmaps, or expand into a repository-wide cloning refactor.
Preparation: Existing injectable HTTPClient/RoundTripper seam and local-server
tests in `internal/adapters/whisperx/http_test.go`; app WhisperX wiring tests.
1. Add a deterministic transport-boundary regression: consume part of the
original request, close it as a transport would, call GetBody, consume and
parse the replay using the request Content-Type, then return valid JSON.
Verify exact audio/language, installed output, and one application attempt.
This must fail before replay support; merely checking non-nil GetBody is
insufficient. Do not assert Go's internal retry count or require a flaky
handcrafted HTTP/2 peer in the default suite.
2. Cover replay-open/read failure, early response, parent cancellation during
replay, and disposal of all body handles. Reuse Stage 5 body coverage rather
than repeat every low-level case. Prove an intentionally abandoned original
body does not turn a valid replay response into failure.
3. Retain local HTTP tests for application retries, non-retryable responses,
response limits, invalid JSON, and bounded streaming; exercise that an
exhausted HTTP failure still leaves earlier speaker checkpoints reusable
using the existing stage/app seams.
4. Record the verification boundary honestly: controlled transport replay proves
Narratio's body contract, not recovery from every real proxy PROTOCOL_ERROR.
Verify: `go test ./internal/adapters/whisperx ./internal/app ./internal/stage`
and race tests for changed packages.
Done when the replay defect is protected at the adapter boundary and the
application retry/checkpoint behavior remains intact.
## Stage 7: Document and Validate the Completed Feature
Status: Complete.
Completion marker (2026-09-22): Updated the canonical transcribe, merge,
manifest, WhisperX, operations, and troubleshooting documentation for durable
speaker reconciliation, serialized checkpoints, aggregate gating, durable
force reset, legacy upgrade and downgrade limits, replayable streaming bodies,
resource ownership, and the absence of an exactly-once remote guarantee.
Reviewed the completed implementation against the feature roadmap and
implementation-discipline guidance. Verified with `go test ./...`,
`go test -race ./...`, `go vet ./...`, `go build ./...`,
`go test ./internal/doccheck`,
`go test ./internal/config -run '^TestExamplesLoadAndValidate$'`, and
`git diff --check`.
1. Update canonical transcribe and manifest internals for the collection,
reconciliation, durable-reset force behavior, serialized saves, and aggregate
gate. Remove obsolete all-speakers-before-canonical-write descriptions.
2. Update the WhisperX contract for replayable bounded streaming, application
attempts versus transport replay, resource ownership, and the lack of an
exactly-once server guarantee. Do not claim the remote protocol issue is fixed.
3. Update operations/troubleshooting with ordinary resume versus force, durable
partial success, old-manifest one-time retranscription on selecting transcribe,
downgrade limitations, and private-model-change guidance. No new CLI flags
or promises of automatic historical-file recovery. Keep detailed wire and
storage contracts in their canonical owners.
4. Review changes against the feature roadmap and the implementation-discipline
guidance. Do not elevate selected storage encoding, internal ordering, test
synchronization, or exact error wording into new product guarantees.
5. Run the full repository gates below; document results and update completion
statuses only after they pass. Leave release preparation for a separate task.
```sh ```sh
go test ./... go test ./...
@@ -422,10 +278,12 @@ go test ./internal/config -run '^TestExamplesLoadAndValidate$'
git diff --check git diff --check
``` ```
Done when duplication is removed, recovery remains covered, documentation
matches the implementation, and all gates pass.
## Open Questions ## Open Questions
None requiring user input. The agreed roadmap supplies the product decisions; None requiring a product decision. A private stage-level helper is selected
this plan selects a completed-only manifest collection, conservative old-record because persistence already has an owner and the duplication is execution
handling, and pull-based replay as implementation choices. No stronger remote plumbing. Model unification, library extraction, generalized transactions, and
reliability guarantee, new operator restriction, or generic orchestration broader cloning cleanup remain out of scope.
surface is proposed.

View File

@@ -15,9 +15,9 @@ peer `PROTOCOL_ERROR` and an inability to replay the request body without
`Request.GetBody`. `Request.GetBody`.
These are related reliability improvements, not a promise to eliminate all These are related reliability improvements, not a promise to eliminate all
service failures. This document owns the target state and scope; the service failures. This document records the target state and scope. The
[implementation plan](implementation.md) owns ordered changes and concrete completed implementation sequence is retained in Git history; current behavior
storage and replay decisions. is documented in the canonical guides linked below.
## Pre-implementation Limitations ## Pre-implementation Limitations