Replay v0 extraction inventory

Status: COMPLETE for current main at 8f468231cb02632b2f2062e501ce7ccdd708675c

Recorded: 2026-07-30

This inventory implements Task 2 of AGENT_HARNESS_OPERATIONS.md. It identifies candidate assets without moving, copying, executing, or opening private transcript/corpus data. Git and the published repository state outrank older branch or handoff prose.

Decision vocabulary

Keep legacy reference — useful evidence in this existing repository, but never copied into replay v0.
Rewrite — retain only the contract or case idea and implement it afresh against the v0 schemas. No source-level copy.
Drop — unrelated, unsafe to extract, or explicitly outside the product boundary.
Create — a required v0 asset does not exist and must be authored from privacy-reviewed synthetic or redacted inputs.

Effort estimates are extraction effort within the operations program, not historical effort or a promise about the unconfirmed launch calendar.

Candidate assets

Current path or source	Current purpose	Disclosure/privacy risk	Decision	Estimated extraction effort
`templates/hooks/dispatch.py` at proposed freeze commit `02bd14c`	Live legacy deny-floor implementation and source of recorded policy identity	Low disclosure risk in the already-public repository, but copying it would violate the v0 no-legacy-runtime boundary and carry repository-specific enforcement semantics	Keep legacy reference; drop from bundle. Identify it only by commit/version in a recording manifest	0.1 h
`templates/hooks/smoke_test.py`	Self-counting canonical allow/deny matrix for the legacy dispatcher	Mostly synthetic commands, but tightly coupled to legacy effects, tiers, runtime responses, and host probes; wholesale reuse would turn v0 into a parser conformance suite	Rewrite selected case concepts. Re-author a small privacy-safe charter; do not copy the harness	0.6 h in Task 9
`tests/test_prefix_wrapper_crossproduct.py`	Large both-direction wrapper/prefix gate over the legacy deny corpus	Synthetic, but contains legacy parser taxonomy, expected bypasses, and executable-shape fixtures far beyond v0	Drop implementation; rewrite only a few distinct regression ideas that survive charter review	0.2 h in Task 9
Focused dispatcher tests under `tests/test_*.py` other than replay tests	Unit/contract evidence for floor parsing, Git/config probes, worktrees, adapters, and false-positive fixes	Low direct disclosure risk; high coupling and scope contamination	Keep legacy reference; drop from bundle. Cite issue/family evidence rather than copying tests	0.1 h
`FLOOR_LIMITATIONS.md` plus public issues #21, #96, #133–#138 and merged PR #132 review evidence	Decision history, measured false-positive families, accepted limits, and unresolved review claims	Public evidence, but examples can still encode legacy-specific assumptions and must not be treated as truth labels automatically	Keep as background; rewrite case rationales. Never infer charter class from a legacy verdict	0.3 h in Task 9
`scripts/replay_corpus.py`	Offline comparison of two dispatcher versions over real Claude/Codex transcript commands; can emit raw JSON/corpus cache	High. Its own privacy contract warns that commands can include repository paths, branch names, private prompts, and pasted tokens; full command text belongs only in scratch output	Keep legacy reference; drop/rewrite. v0 gets new schema-bound sources, comparison, and reports with no transcript scanner	0 h copy; Tasks 4–8 implement replacements
`tests/test_replay_corpus.py`	Tests the legacy transcript scanner, multiprocessing, offline stubs, cache integrity, and old/new dispatcher invocation	Medium: fixtures and assumptions include machine/repository identity signals; behavior is coupled to private transcript ingestion and legacy `check()` signatures	Drop/rewrite. Create focused v0 unit and contract tests from the operations contract	0.4 h across Tasks 3–8
`tests/fixtures/floor_1_2_0_dispatch.py`	Historical dispatcher source fixture used to exercise legacy replay signature compatibility	Low direct disclosure risk, but it is legacy implementation code rather than a decision recording	Keep legacy reference; drop from bundle. Recorded `PolicyDecision` JSONL replaces executable baselines	0 h
`scripts/generate_curl_option_arity_fixture.py`, `tests/fixtures/curl_8_21_0_option_arity.json`, `tests/test_curl_option_arity.py`	Generates and verifies Curl option-arity data for the legacy parser	Low-to-medium: includes external URL/path vocabulary and is unrelated to decision replay	Drop. No v0 contract depends on Curl option parsing	0 h
`HANDOFF.md` and historical branch/PR narratives	Operational history for prior floor work	Stale by construction; may cite local scratch paths and superseded heads	Drop. Freeze record and live Git/GitHub evidence are the extraction sources of truth	0 h
Occupied `.worktrees/replay-tool` checkout on `tooling/corpus-replay`	Local continuation of the already-merged legacy replay tool	High work-loss/collision risk. It is about 310 commits behind current `main` and has a staged, uncommitted `scripts/replay_corpus.py` change (+209/−9) owned by another writer	Do not touch or extract. Current `main` plus this inventory define the usable baseline	0 h
Tracked JSONL corpus or `PolicyDecision` recording	Required v0 event/decision input	Absent. `git ls-files` found no tracked `*.jsonl`, `replay_v0/`, recorded baseline, or charter corpus	Create only after schemas exist, using reviewed synthetic/redacted records and exact-byte manifests	0.7 h in Tasks 4 and 9

Baseline recording gap

There is no committed decision recording that satisfies PolicyDecision v1. In particular:

the historical dispatcher fixture is executable Python, not recorded decisions;
the smoke matrix embeds expected legacy outcomes in test code, not stable event-keyed JSONL;
legacy replay JSON/cache output is scratch-only and can contain unreviewed private commands; and
no manifest binds a committed decision set to a legacy commit, environment note, or digest.

Task 4 must therefore implement a recorded-decision source against synthetic fixtures. Task 9 must author or owner-review a matching privacy-safe baseline recording and sidecar manifest. Missing, duplicate, malformed, or digest-mismatched records must remain indeterminate; no task may fill the gap by executing the frozen dispatcher in the public product repository.

Safe inputs for the charter

The charter may draw ideas, not files, from these reviewed categories:

canonical destructive Git and filesystem shapes already represented with fictional identifiers;
benign quoted documentation, help, dry-run, and non-executing near-misses;
a small number of historically important wrapper, redirection, interpreter, and worktree cases; and
deliberately opaque cases whose expected class is opaque, independent of any legacy decision.

Every selected case must be re-authored with a new stable event_id, fictional/redacted cwd, approved provenance, and no owner path, private URL, repository identity, credential-like value, customer/staff/student data, or raw transcript text. A case that cannot preserve its relevant structure after sanitisation remains private and is replaced with a synthetic case.

Explicit extraction denylist

Task 10’s allowlisted public bundle must exclude:

templates/hooks/ and every legacy dispatcher fixture;
scripts/replay_corpus.py and tests/test_replay_corpus.py;
all raw transcript, cache, report, or locally generated replay output;
.git/, Git history, .worktrees/, local failure ledgers, and machine state;
HANDOFF.md, private scratch references, and unreviewed historical examples; and
any file not explicitly named by docs/extraction/public-v0-manifest.json.

Prerequisites and blockers carried forward

The owner confirmed on 2026-07-30 that the live GitHub repository remains public for now. Treat every tracked artifact as immediately public; H-12 is complete.
At the recorded base, CLAUDE_CONFIG_OPERATIONS.md and REPLAY_TOOL_PRODUCT.md were absent from the searched repository, history, and local authority locations. Subsequent main commit a35ff70 added REPLAY_TOOL_PRODUCT.md, and 6d6e22e added CLAUDE_CONFIG_OPERATIONS.md; H-9 is complete.
Ruff and Black remain pinned, Pytest 9.0.3 is now an approved pinned development dependency, and the owner approved both the declared Pytest compatibility commands and authoritative dependency-free unittest CI lanes; H-10 is complete.
The immutable tag, launch calendar, and time allocation were owner-confirmed on 2026-07-30; H-8 and H-11 are complete.

Final name, licence, clean-repository creation, and release remain separate product gates. They do not block the approved local extraction implementation or its two test-runner interfaces.

Verification performed

Inspected tracked-path inventory on current main and the occupied replay checkout’s Git metadata without modifying it.
Confirmed PR #33 is merged and no current PR owns tooling/corpus-replay.
Used static help/docstring and symbol searches; did not run replay against transcripts or open raw corpus/cache output.
Confirmed no tracked replay_v0/, JSONL corpus, or recorded decision baseline exists.
Confirmed this task moved or copied no source asset.