Purpose: Preserved candidate product brief for a possible future public decision-replay extraction. Status: ACTIVE for the replay tool since 2026-09-29 (RP-006, AGENT_HARNESS_OPERATIONS AH-006): the owner approved the clean repository, name, licence and publication. Outreach still needs owner direction. Text below that says DEFERRED records the earlier plan. Authority relationship: AGENT_HARNESS_AGENT_BRIEF.md wins for current repository mission and defers public extraction until internal stability and demonstrated demand. If AH-10 is later unlocked, this document supplies candidate product requirements; AGENT_HARNESS_OPERATIONS.md still wins for replay-kernel contracts and CLAUDE_CONFIG_OPERATIONS.md for owner supervision. Last-reviewed date: 2026-07-30 Owner: Cristian Tcaci

Decision register

• HISTORICAL RP-001: The owner removed the former calendar launch deadline.
• HISTORICAL RP-002: The original public-product programme recorded 24 focused hours: 13 hours for replay extraction and 11 hours for this product brief. That accounting is not a global workbench cap.
• HISTORICAL RP-003: A 2026-09-30 continuation review was recorded for the former launch plan; the date does not trigger extraction.
• RP-004 (superseded by RP-006): Final repository, package, and CLI naming is an AH-10 decision, not the next owner question. CharterReplay / charter-replay remains only a candidate.
• RP-005 (superseded by RP-006): Licensing is an AH-10 decision after provenance review, not the next owner question. No licence recommendation authorises a change.
• DECISION RP-006 (2026-09-29, owner in chat): RP-004 and RP-005 are decided: the repository, package and CLI are charter-replay, licensed Apache-2.0, published from Chris0Jeky/charter-replay. See AGENT_HARNESS_OPERATIONS AH-006.

Deferred repository design

• DEFERRED: A future public replay product may use a clean new repository rather than a public release branch inside agent-harness; AH-10 must re-evaluate and explicitly approve that choice. Do not create it now.
• DECISION: The clean repository receives only the allowlisted replay kernel, schemas, synthetic/redacted charter corpus, deterministic tests, product documentation, and examples.
• CONSTRAINT: The clean repository receives no legacy dispatcher source, private configuration, private corpus, raw history, machine paths, client/project identity, or private Git history.
• CONSTRAINT: agent-harness remains the active workbench. Its internal replay module is the Policy Lab and evidence centre alongside Doctor, estate, adapter, measurement, and bounded-guard domains.
• DEFERRED: If extraction is later approved, clean copying remains the candidate because an independent product should not require the legacy environment and history rewriting is not an acceptable shortcut.
• CONSTRAINT: Repository creation, visibility selection, public push, package-name reservation, and release publication require owner review under CLAUDE_CONFIG_OPERATIONS.md#autonomy-boundary.

Deferred candidate names

• CONSTRAINT: Before public creation, the owner must recheck GitHub, PyPI, package-manager, and basic web-name availability. Search findings are time-sensitive and are not a reservation.
• CONSTRAINT: The current internal module remains replay_v0. The provisional public names below apply only if AH-10 later unlocks extraction.

Product definition

• DECISION: Product sentence:

CharterReplay compares two coding-agent command-policy decision sources over a pinned corpus and fails when the candidate introduces configured safety or interruption regressions.

• DECISION: The primary user is a developer or maintainer changing a command guard, hook, or policy who needs a reproducible before/after decision report.
• DECISION: The primary portfolio outcome is a credible engineering case study backed by executable proof.
• DECISION: External product adoption is conditional and is evaluated separately at the confirmed continuation review on 2026-09-30.
• CONSTRAINT: The product is described as decision replay, not agent replay, environment replay, command execution, sandboxing, or complete safety enforcement.
• CONSTRAINT: The public core implements exactly the technical contract in AGENT_HARNESS_OPERATIONS.md; this document does not redefine the event or decision schemas.

Non-Goals

• CONSTRAINT: No live command blocking, approval workflow, sandbox, policy authoring language, shell interpreter, generic tool-event model, agent trajectory replay, observability platform, configuration doctor, promotion engine, or blueprint plugin.
• CONSTRAINT: No claim that a policy is universally safe, that a command was executed, or that an original failure was reproduced.
• CONSTRAINT: No hosted service, account system, database, web UI, telemetry collection, or network requirement in v0.
• CONSTRAINT: No broad integration framework. External systems use the generic process contract or recorded decisions.
• CONSTRAINT: No benchmark, leaderboard, academic paper, or “industry standard” language for the initial charter corpus.

Sixty-second README specification

• DECISION: The first visible screen of README.md contains exactly these five product elements before architecture or history.

1. One-line purpose

• DECISION: Use:

Compare coding-agent command policies against a pinned replay corpus and catch decision regressions before release.

2. Demo GIF

• DECISION: The GIF storyboard is 60–90 seconds and shows:
  1. A single charter-replay replay command.
  2. A baseline recording and candidate process policy.
  3. A summary with one newly allowed dangerous case, zero newly denied benign cases, and one indeterminate case.
  4. Exit code 1 and the generated Markdown report.
  5. The limitation line: “Recorded decisions were compared; no command was executed.”
• CONSTRAINT: The GIF contains fictional paths, synthetic commands, no username, no home directory, no private repository, and no terminal history outside the demo.

3. One install command

• DECISION: The release README shows one install command only.
• OPEN: Preferred command, if the PyPI name is secured:

pipx install charter-replay

• OPEN: If PyPI publication is not approved or the name is unavailable, replace the block with one exact immutable Git-tag install command. Do not show both alternatives above the fold.

4. One replay command

• DECISION: Use this shape with paths included in the repository:

charter-replay replay \
  --baseline recorded:examples/legacy-baseline/decisions.jsonl \
  --candidate process:python,examples/candidate-policy/policy.py \
  --corpus corpora/charter/events.jsonl \
  --output reports/demo

5. One limitation statement

• DECISION: Use:

This re-evaluates recorded command events and decisions. It does not reproduce the original agent, environment, shell effects, or operating-system boundary.

• CONSTRAINT: The README may link to the postmortem, threat model, schemas, and extension guide below the first screen, but it may not insert feature tables or architecture diagrams before the five required elements.

Public repository layout

• DECISION: The clean repository uses this bounded layout:

charter-replay/
  REPLAY_TOOL_PRODUCT.md
  README.md
  LICENSE
  SECURITY.md
  CONTRIBUTING.md
  pyproject.toml
  src/charter_replay/
    __init__.py
    cli.py
    compare.py
    corpus.py
    digests.py
    manifests.py
    policy_sources.py
    reports.py
    schemas/
  corpora/charter/
  examples/
    legacy-baseline/
    candidate-policy/
    dcg/
  tests/
    unit/
    contract/
    charter/
  docs/
    decision-replay-boundary.md
    postmortem.md
    continuation-review.md
  .github/workflows/

• CONSTRAINT: No legacy/, experimental/, private overlay, raw transcript, or copied planning archive is included.
• CONSTRAINT: Background design is summarised through the postmortem; the public repository does not ship the entire planning corpus.

Versioning and release process

• DECISION: Product releases use Semantic Versioning; the first public release is v0.1.0.
• DECISION: JSON schema identifiers version independently through names such as command-event.v1; a package patch release may fix implementation without changing a schema identifier.
• CONSTRAINT: A breaking schema or CLI change before 1.0.0 still requires a release-note callout, migration note, and updated fixtures.
• CONSTRAINT: Every release candidate must pass all quality gates, install into a clean environment, run the README command, and reproduce the checked-in demo report semantics.
• CONSTRAINT: Tags are annotated and created only after owner review. Public push, GitHub release, and package publication are human actions recorded in HUMAN_TODO.md.
• CONSTRAINT: Release artifacts contain source and wheel only; no binary executable, private corpus, generated local report containing machine data, or vendored third-party guard.
• CONSTRAINT: The package has no runtime dependency unless the owner approves one. Development dependencies are declared in one owner-reviewed dev extra.
• DECISION: Release checklist:
  1. Update version and changelog.
  2. Run fast, lint, full tests, charter replay, and build.
  3. Install the wheel into a clean virtual environment.
  4. Run the README replay command.
  5. Review generated files for private paths and secrets.
  6. Owner approves licence, tag, public push, and publication.
  7. Create v0.1.0, publish release notes, and verify installation.

Licence decision

• OPEN: Final licence is owner-selected after checking the existing agent-harness licence and the provenance of every copied file.
• DECISION: Recommended default is Apache-2.0 because it is permissive, includes an explicit patent grant, and is legible to organisational users.
• CONSTRAINT: If any extracted file is incompatible with Apache-2.0, it is excluded or the repository uses the compatible licence chosen by the owner. An agent must not relicense copied code by assumption.
• CONSTRAINT: LICENSE, licence headers, pyproject.toml, README badges, and contribution terms must agree before release.

SECURITY.md contract

• DECISION: SECURITY.md must state supported versions, private vulnerability-reporting route, response expectations without guaranteed SLA, security boundary, privacy expectations for corpus submissions, and safe reproduction rules.
• CONSTRAINT: It must explicitly state that the tool does not execute corpus commands and is not a sandbox or enforcement layer.
• CONSTRAINT: It must instruct reporters not to submit credentials, private repositories, raw transcripts, customer data, or operational exploit payloads when a synthetic reproduction is possible.
• CONSTRAINT: GitHub private vulnerability reporting is preferred if enabled; the owner must supply any public security contact.
• CONSTRAINT: Security-policy edits require owner review.

CONTRIBUTING.md contract

• DECISION: CONTRIBUTING.md requires a scoped issue or proposal before new features, one-purpose PRs, synthetic fixtures, schema validation, all quality gates, and an explicit non-goal check.
• CONSTRAINT: Contributors may add a corpus case only with provenance, privacy review, case class, rationale, and a demonstration that it is not a duplicate.
• CONSTRAINT: New runtime dependencies, new policy effects, schema fields, integrations, or public claims require an owner-approved design issue.
• CONSTRAINT: Contributions must not include real secrets, private commands, malware, pirated software, destructive payload execution, or instructions for bypassing third-party controls.

DCG compatibility worked example

• DECISION: The first external example targets Destructive Command Guard (DCG) because it is a direct command-guard neighbour and can be represented through the generic process or recorded-decision contract without changing the core.
• DECISION: The public example compares a pinned DCG decision recording with a candidate recording or an owner-approved local DCG shim over the same charter events.
• CONSTRAINT: The core package contains no DCG-specific import and does not vendor DCG.
• CONSTRAINT: CI reproduces the example from recorded decisions. Live DCG installation and capture are owner-reviewed optional steps documented separately.
• CONSTRAINT: The report is descriptive and does not rank or declare a universal winner.
• CONSTRAINT: The example records DCG version/commit, recording digest, corpus digest, unsupported cases, and the exact reproduction command.

Outreach checklist

• DECISION: Prepare five separate, concise messages after v0.1.0 is reproducible. The owner sends them; agents may draft only.
• CONSTRAINT: Each message includes one report link, one reproduction command, one limitation sentence, and one specific question. No bulk generic announcement is sent as direct outreach.

• CONSTRAINT: If a named project or maintainer has changed before launch, the owner revalidates the target. Agents must not guess contact identities.

Success and continuation gates

Portfolio completion

• DECISION: Portfolio completion is achieved when all of these exist:
  • A published or owner-approved postmortem.
  • A tagged v0.1.0 clean repository.
  • One reproducible replay command.
  • A small charter corpus.
  • A 60–90 second demo.
  • Clear limitations.
  • Five drafted targeted messages.
• DECISION: Portfolio completion does not depend on stars, page views, or external adoption.

Product continuation

• DECISION: The continuation review remains 2026-09-30; it does not delay v0 implementation.
• DECISION: Continue expanding the public product only if at least one strong signal exists:
  • One external guard maintainer uses, links, contributes to, or requests the workflow.
  • Three unaffiliated users successfully run it and provide concrete policy or corpus feedback.
  • One external integration or corpus contribution is opened.
  • A real user identifies a policy regression they would not otherwise have caught.
• CONSTRAINT: Stars, impressions, likes, and post views are contextual signals only; they do not satisfy the continuation gate by themselves.
• DECISION: If no strong signal exists at review, freeze the tool at a clean v0.x, fix only serious defects, retain the postmortem and corpus, and return engineering time to higher-priority work.
• CONSTRAINT: Failure to meet the continuation gate does not reopen the legacy parser or authorise a broader product pivot.

Deferred backlog with evidence triggers

• DEFERRED: Every row below is deferred exactly as stated. No row is part of v0.

• CONSTRAINT: A deferred item unlocks only when its named trigger is evidenced in an issue and the owner approves a new bounded task contract.

Autonomy and self-improvement references

• CONSTRAINT: Apply the owner-review and destructive-operation rules from CLAUDE_CONFIG_OPERATIONS.md#autonomy-boundary.
• CONSTRAINT: Use the failure-ledger schema from CLAUDE_CONFIG_OPERATIONS.md#failure-and-friction-ledger; the recommended untracked path is .local/failure-ledger.jsonl.
• CONSTRAINT: Use the manual promotion template from CLAUDE_CONFIG_OPERATIONS.md#manual-self-improvement-loop; no automatic skill, rule, documentation, or integration creation is permitted.
• CONSTRAINT: Any discovered work outside this product scope is appended to HUMAN_TODO.md as a backlog proposal with a named evidence trigger; it is not implemented under v0.
• CONSTRAINT: Public contributors follow CONTRIBUTING.md; the private operating contract governs owner-supervised agents and is not a substitute for contributor terms.

Quality gates

• DECISION: Pytest 9.0.3 is approved for the extraction compatibility gate. Ruff is already pinned in the extraction repository. Adding or upgrading build, or changing these versions in the clean product repository, still requires owner review.
• DECISION: Required local commands are:

python -m pytest -q tests/unit tests/contract
python -m ruff check .
python -m pytest -q
python -m build
charter-replay replay --baseline recorded:examples/legacy-baseline/decisions.jsonl --candidate process:python,examples/candidate-policy/policy.py --corpus corpora/charter/events.jsonl --output .local/release-proof

• CONSTRAINT: The first command is the fast lane and must complete in under 60 seconds without network access.
• CONSTRAINT: Required CI checks before merge are fast, lint, tests, charter-replay, build, privacy-scan, and operations-contract.
• CONSTRAINT: privacy-scan checks tracked files and built artifacts for forbidden owner paths, private repository names, credential patterns, and files outside the extraction allowlist; it does not upload repository content.
• CONSTRAINT: CI has read-only repository permissions unless a separately reviewed release job requires more.

Historical public-product allocation

• HISTORICAL: The ten tasks below carried an 11-hour allocation in the former public launch plan. Together with the 13-hour replay extraction allocation, that produced the historical 24-hour figure.
• CONSTRAINT: No task below is active. The figure neither caps the wider workbench mission nor authorises clean-repository creation, naming, licensing, release, or outreach.

Deferred AH-10 task candidates

• CONSTRAINT: Every task below is locked until AH-10 records internal stability, demonstrated demand, duplicate-search results, current provenance/privacy evidence, and the required owner decisions. Listing a task is not authorisation.

Task 1 — Create the clean local repository and reserve the working identity

• Classification: DECISION
• Time box: 1 hour
• Sequence: Deferred until AH-10 is explicitly unlocked; PR #140 closing does not unlock it.
• Objective: Create a local clean repository using the provisional charter-replay name and add only root governance files.
• Paths touched: New repository root, REPLAY_TOOL_PRODUCT.md, .gitignore, HUMAN_TODO.md.
• Acceptance criteria: git log begins with clean public-intent history; no file from private Git history is imported; HUMAN_TODO.md contains final name, visibility, licence, and remote-creation actions.
• Verify: git status --short && git log --oneline --max-count=3 && python -c "from pathlib import Path; assert Path('REPLAY_TOOL_PRODUCT.md').is_file()"
• Out of scope: Creating a public remote, reserving PyPI, adding product source, selecting licence without owner review, or copying planning archives.
• Stop condition: Halt before any remote creation or public push and when the final name is not owner-approved.

Task 2 — Import the allowlisted replay bundle and package skeleton

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after Task 1 and exact-bundle provenance review.
• Objective: Copy the approved v0 bundle, place it under src/charter_replay, and create the minimal package metadata.
• Paths touched: src/charter_replay/, tests/, corpora/, examples/legacy-baseline/, examples/candidate-policy/, pyproject.toml, tools/product_gate.py.
• Acceptance criteria: Only manifest-listed files are copied; import succeeds; provisional CLI help works; no runtime dependency is added; provenance is recorded in the initial commit message or docs/extraction-provenance.md; a small standard-library product_gate.py exposes self-test, readme, public-docs, demo, privacy, outreach, release, and release-handoff checks without becoming a general framework.
• Verify: python tools/product_gate.py self-test && python -m pytest -q tests/unit tests/contract
• Out of scope: Legacy source, private tests, history import, DCG files, release publication, or source redesign.
• Stop condition: Halt when any source file is absent from the approved extraction manifest, has unclear licence provenance, or contains private identifiers.

Task 3 — Build the sixty-second README

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after Task 2 and final package/install identity selection.
• Objective: Implement the exact first-screen README specification and a concise limitations-first continuation.
• Paths touched: README.md, static placeholder path for the demo GIF.
• Acceptance criteria: The first screen contains one purpose line, one GIF placeholder, one install command block, one replay command block, and one limitation statement; no architecture diagram or feature matrix precedes them.
• Verify: python tools/product_gate.py readme
• Out of scope: Long research summary, competitor comparison, benchmark claims, roadmap expansion, or multiple install methods above the fold.
• Stop condition: Halt if the final package name or install source is unresolved; leave one marked placeholder and add RP-004 to HUMAN_TODO.md rather than showing multiple commands.

Task 4 — Add licence-decision package, security policy, and contribution contract

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after Task 2; provenance work may run alongside Task 3.
• Objective: Prepare SECURITY.md, CONTRIBUTING.md, provenance evidence, and the owner licence decision item.
• Paths touched: SECURITY.md, CONTRIBUTING.md, docs/licence-provenance.md, HUMAN_TODO.md; LICENSE only after owner selection.
• Acceptance criteria: Both documents meet their contracts; provenance lists every extracted source path and prior licence; no unsupported contact or SLA is invented.
• Verify: python tools/product_gate.py public-docs SECURITY.md CONTRIBUTING.md docs/licence-provenance.md
• Out of scope: Selecting or changing the licence autonomously, adding a code of conduct without owner direction, or copying third-party policy text.
• Stop condition: Halt before writing LICENSE when provenance is incomplete or licence compatibility is uncertain.

Task 5 — Record the demo GIF from a fictional clean terminal

• Classification: DECISION
• Time box: 1.5 hours
• Sequence: Start after Tasks 3 and 4 establish the public documentation boundary.
• Objective: Produce the 60–90 second storyboard asset showing a configured regression and generated report.
• Paths touched: docs/assets/demo.cast or source recording, docs/assets/demo.gif, README.md.
• Acceptance criteria: The demo follows the five storyboard steps; all paths and commands are fictional; it shows exit 1; no unrelated terminal content is visible; the GIF remains legible at README width.
• Verify: python tools/product_gate.py demo docs/assets/demo.gif && python tools/product_gate.py readme
• Out of scope: Voice-over, promotional video, multiple scenarios, live destructive commands, or private screen recording.
• Stop condition: Halt and discard the recording if any username, token, private path, notification, browser tab, or unrelated shell history appears.

Task 6 — Add release CI and build a local v0.1.0 candidate

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after the source, documentation, and demo inputs are stable.
• Objective: Implement least-privilege CI and produce a locally installable release candidate without publishing it.
• Paths touched: .github/workflows/ci.yml, optional owner-reviewed release workflow, pyproject.toml, CHANGELOG.md.
• Acceptance criteria: Required checks are present; workflow permissions default to read-only; wheel and source archive build; clean-environment install and README replay succeed.
• Verify: python -m ruff check . && python -m pytest -q && python -m build && python tools/product_gate.py release dist/
• Out of scope: Trusted publishing setup, PyPI upload, GitHub release, signing infrastructure, or automatic tag creation.
• Stop condition: Halt if a workflow requests write, id-token, package, or release permissions without a separately reviewed release need.

Task 7 — Finish the postmortem and launch bundle

• Classification: DECISION
• Time box: 1.5 hours
• Sequence: Start after executable proof exists; it may proceed alongside Task 8.
• Objective: Turn the existing retrospective into a concise public postmortem linked to executable proof.
• Paths touched: docs/postmortem.md, docs/launch-note.md, README link.
• Acceptance criteria: The postmortem states the original hypothesis, measured 12–14% false-positive/friction result if supported by retained evidence, review-fix loop, architecture correction, preserved assets, decision-replay boundary, and explicit limitations; every quantitative claim has an internal evidence reference.
• Verify: python tools/product_gate.py public-docs docs/postmortem.md docs/launch-note.md && python tools/product_gate.py privacy docs/postmortem.md docs/launch-note.md
• Out of scope: Research literature review, blaming specific agents or vendors, broad market claims, or publishing unsupported internal metrics.
• Stop condition: Halt on any number that cannot be traced to retained evidence; replace it with a qualitative statement or request owner confirmation.

Task 8 — Produce the DCG compatibility worked example

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after Task 2 when pinned, licence-compatible source material is available.
• Objective: Add a thin example that compares pinned DCG decision recordings or an owner-approved local shim without changing core code.
• Paths touched: examples/dcg/README.md, examples/dcg/decisions-*.jsonl, manifests, examples/dcg/report/, optional examples/dcg/shim.py.
• Acceptance criteria: The example names the pinned source version, unsupported cases, digests, exact replay command, and neutral findings; CI reproduces it from recordings without installing DCG.
• Verify: charter-replay replay --baseline recorded:examples/dcg/decisions-baseline.jsonl --candidate recorded:examples/dcg/decisions-candidate.jsonl --corpus corpora/charter/events.jsonl --output examples/dcg/report
• Out of scope: Ranking DCG, bundling its binary/source, installing it without approval, opening a maintainer PR, or adding a DCG-specific core adapter.
• Stop condition: Halt if live capture requires a download, dependency, network access, or uncertain CLI invocation; prepare the recording contract and route capture to HUMAN_TODO.md.

Task 9 — Draft the five targeted outreach messages

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after the release candidate and launch evidence are stable.
• Objective: Draft one tailored message for each named target and one public launch post, without sending any message.
• Paths touched: docs/outreach/01-dcg.md through 05-governance-toolkit.md, docs/outreach/public-launch.md, HUMAN_TODO.md.
• Acceptance criteria: Each direct message contains one report, one command, one limitation, and one specific question; the public post links the postmortem, demo, and repository; no message claims endorsement.
• Verify: python tools/product_gate.py outreach docs/outreach
• Out of scope: Sending messages, tagging maintainers publicly without approval, repeated cross-posting, automated outreach, or follower scraping.
• Stop condition: Halt if a named handle cannot be verified at launch; leave the draft unsent and add a human verification item.

Task 10 — Prepare the owner release handoff and continuation record

• Classification: DECISION
• Time box: 1 hour
• Sequence: Start after Tasks 1 through 9 are complete or factually parked.
• Objective: Present one owner-review packet for name, licence, remote, tag, publication, outreach, and the 2026-09-30 continuation review.
• Paths touched: docs/release-handoff-v0.1.0.md, docs/continuation-review.md, HUMAN_TODO.md.
• Acceptance criteria: All quality-gate outputs and reproduction commands are linked; cumulative hours are recorded; every human-only release action is explicit; continuation criteria are prefilled without fabricated results.
• Verify: python tools/product_gate.py release-handoff docs/release-handoff-v0.1.0.md docs/continuation-review.md && git status --short
• Out of scope: Public push, tag, PyPI publication, outreach sending, changing continuation criteria after results are known, or consuming more than the remaining hour budget.
• Stop condition: Halt at the owner-review boundary. No public action occurs until the owner resolves RP-004 and RP-005 and approves the release packet.