continuity-handoff

Files (6)

• SKILL.md
• openai.yaml · agents/
• packet-template.md · references/
• recall-codex-history.mjs · scripts/
• validate-handoff-acceptance.mjs · scripts/
• write-continuity-packet.mjs · scripts/

SKILL.md (first lines)

# Continuity Handoff

Continue work in a fresh task without duplicating the legacy transcript.

The default path is fast, packet-only, and durable. Historical recall is
exceptional, question-driven, read-only, and bounded.

## Modes

- **Create**: produce a continuity packet for a fresh task.
- **Acknowledge**: verify a received packet before implementation.
- **Recall**: recover one missing historical fact from the legacy Codex rollout.

Do not use native task fork as a desbloat mechanism unless a fresh measurement
proves that the destination does not inherit material transcript weight.

## Create

1. Read the current repo instructions and current task state.
2. Identify the source task by Work, human title, Thread ID, Profile, Profile
   Placement, Device, and Project Binding. Record the Codex Project Root,
   Canonical Repo Root, and Operational CWD separately. Never identify the task
   by ID alone or assume that the saved project root is itself a Git repository.
3. Capture the current goal or explicitly state that no active goal exists.
4. Reference existing commits, specs, ledgers, issues, and proof artefacts. Do
   not paste their full contents.
5. Record accepted decisions, rejected directions with reasons, protected
   state, open gates, known historical gaps, and the exact resume point.
6. Resolve the source `CODEX_HOME` and `rollout_path` while Codex remains open.
   Record these under Device Observations, never as portable Work identity. The
   bundled reader admits the selected live SQLite database from its header and
   existing sidecars, then performs a short read-only exact-Thread-row lookup
   in a disposable child process with a wall-clock termination deadline. It
   closes the connection and child immediately. It does not copy the database. The reader's file
   allowlist is the selected `state_5.sqlite`, its existing WAL coordination
   sidecars, `session_index.jsonl`, and the resolved rollout only.
7. Fill [the packet template](references/packet-template.md) completely,
   including goal, checkpoint, accepted decisions, failures not to repeat,
   references, inherited usage, resume point, and Evidence Lineage.
8. Scan and write the completed draft with the bundled writer. In a Git project
   the default final path is `.traverse/continuity/<work-slug>.md`:

```bash
node scripts/write-continuity-packet.mjs \
  --packet-file <completed-draft.md> \
  --project-root <verified-project-root> \
  --work-slug <stable-work-slug>
```

   The writer checks high-confidence secrets before writing, refuses ignored
   destinations, writes atomically beside the destination, and verifies that
   the packet is tracked or visible in `git status`. If a secret is detected,
   stop with `packet_secret_detected` and ask the operator what to do. Never
   redact, rewrite, commit, or create timestamp copies automatically.
9. If the operator explicitly asks to create a new task, deliver the durable
   packet to a fresh native task. A cross-Profile handoff must create a Target
   Thread with a new Thread ID; do not fork or rename the Source Thread.
10. Require acknowledgement before implementation.

The packet must remain compact. A large transcript summary is a failed handoff.

## Acknowledge

Before changing anything, state:

1. the mission and resume checkpoint;
2. accepted facts versus unverified claims;
3. the next read-only action;
4. protected state that will remain untouched;
5. whether implementation has started.

Live repo and runtime truth outrank the packet. Stop and report contradictions.

## Recall

Use recall only when one explicit question cannot be answered by the packet,
current repo, referenced canon, or current runtime evidence.

1. State the unanswered question.
2. Derive a narrow literal search term and optional ISO date.
3. Run the bundled reader:

```bash
node scripts/recall-codex-history.mjs \
  --codex-home <recorded-source-code-home> \
  --thread-id <source-task-id> \
  --rollout-path <recorded-source-rollout-path> \
  --work <recorded-portable-work-identity> \
  --profile <recorded-source-profile> \
  --device <recorded-source-device> \
  --project <recorded-source-project> \
  --query <literal-term> \
  --date YYYY-MM-DD \
  --match-role user \
  --before 1 \
  --after 5 \
  --max-matches 3
```

Run the script from this skill directory while Codex remains open. It opens the
SQLite header before opening it. WAL header mode requires both existing,
readable `-wal` and `-shm` sidecars; the reader fails closed when either or both
are absent and never opens first to create them. The disposable lookup process
opens read-only with a bounded busy timeout, enables `query_only`, reads only
the exact Thread row, and is terminated if the strict wall-clock deadline
expires. The parent validates the rollout identity, streams JSONL verbatim, and
caps output. `-shm` is coordination state, not immutable application content.
Historical evidence is not rewritten or censored. The reader never writes
logical Codex source state, never copies the database, and never requires the
operator to close the app.

Use the exact recorded Device Observation even when the Target Thread is in a
different Profile on the same Device. A primary Target may therefore query an
exact secondary Source by passing the secondary `CODEX_HOME`. If that exact
database, Thread row, or rollout is unavailable, the reader returns
`recall_source_unavailable`. Stop there. Never scan sibling homes, nearby
databases, indexes, or rollout directories to guess a replacement source.

Output schema v2 reports logical source mutation, the bounded SQLite read

From TheAngryPit/TheAngrySkills, MIT.

More from TheAngryPit/TheAngrySkills