| name | continue-from-accepted |
| description | Continue iterative work from the last ACCEPTED result plus only the new requested change, instead of replaying the conversation. Use during long multi-turn editing of a document, config, or code file; when the user says "keep going from what we had", asks for a follow-up change to something already approved, or is near plan/token limits in an iterative session. Saves one integrity-checked accepted result, builds a minimal continuation prompt, and never carries rejected drafts or transcript history forward. |
Continue from accepted
Carry forward the result, not the history that produced it. After a user
accepts a version, later changes need only (accepted result + new change) —
not the transcript, not rejected drafts, not the reasoning that got there.
State rules (read first)
- The state file is EPHEMERAL working state, not a durable store. Keep it
under the OS temp directory:
<tmp>/continue-from-accepted/<task-slug>.json
(resolve <tmp> via python3 -c "import tempfile;print(tempfile.gettempdir())").
Do NOT create per-repo state directories. Durable facts belong in the
memory system; the deliverable itself belongs in the repo. This skill's
state is a session-scoped snapshot, nothing more.
- Only acceptance promotes. Save when the user accepts a version
(explicitly, or by asking for a further change that keeps the rest).
A rejected result is never saved. No accepted result yet → work normally.
- One current result. Each new acceptance replaces the previous state.
No history accumulates.
- No secrets in state. If the accepted result contains credentials,
do not save it; continue without this skill.
Workflow
Resolve <skill-dir> from where this skill is installed. Do this yourself —
never ask the human to run these commands, restate prior work, or manage the
state file.
1. On acceptance — save:
python3 <skill-dir>/scripts/state_delta.py save \
--state <tmp>/continue-from-accepted/<task-slug>.json \
--accepted-file <path-to-accepted-result>
2. On the next change — build the continuation prompt:
python3 <skill-dir>/scripts/state_delta.py packet \
--state <tmp>/continue-from-accepted/<task-slug>.json \
--change "<the user's new requested change>" \
--max-packet-bytes 16000 \
--output <tmp>/continue-from-accepted/<task-slug>-packet.txt
Use the packet as the working input (e.g. as a fresh subagent's prompt)
instead of re-sending conversation history. The packet instructs the worker:
continue from the accepted result, apply only the new change, return the
updated result only.
3. Check without exposing content: state_delta.py inspect --state <file>
prints byte count + sha256 only.
Failure handling
- The state file is sha256-integrity-checked on load; a hash mismatch means
the file was modified outside
save — discard it and re-save from the
actual accepted artifact. Do not hand-edit state files.
- If
packet refuses because the result exceeds the byte cap: select only
the relevant passages of the accepted result for this change (search
tooling), don't raise the cap reflexively.
- Never retry a token-limit or usage-limit failure unchanged. Reduce the
input, pick a cheaper path, or stop and say so.
Provenance
scripts/state_delta.py extracted UNCHANGED from the token-saver skill
(Unlock AI / Ringer, 2026) — reviewed: pure stdlib, no network, no telemetry.
sha256: 0c2d953c5712cbf5d125cbb1dbdbf29eb851f067b783fc9f8b897729858b0c7b
The original's per-repo .token-saver/ state convention was dropped
deliberately: it conflicts with the fleet rule against inventing new state
stores. Temp-dir state is this skill's policy, set here, not in the script.