continue-from-accepted

Étoiles8

Forks1

Mis à jour30 juillet 2026 à 11:27

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.

Installation

Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.

Source

drewid74

drewid74/ai_skills

Ouvrir le dépôt GitHub Voir les dépôts du créateur

Téléchargement

Métiers associésSOC

Basé sur la classification professionnelle SOC

Développeurs de logicielsProfessions informatiques et mathématiques·SOC 15-1252

Explorateur de fichiers

2 fichiers

SKILL.md

readonly

name

continue-from-accepted

description

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.

Plus depuis ce dépôt

même dépôt

quality-test-engineer

drewid74/ai_skills

Use this when: my tests keep failing randomly, write tests for this code, tests pass locally but fail in CI, how do I mock an HTTP call, my test suite is too slow, debug a failing test, increase test coverage, how do I test a database query, write e2e tests for a user journey, run load tests against my API, set up a test framework, add TDD to my workflow, pytest, Vitest, evaluate my agent, LLM eval, agent flaky, RAG eval, prompt regression, hallucination test, Pass@N, LLM-as-Judge, trace agent execution, test my LangGraph agent, agent guardrails, eval dataset, mock agent tools, agent cost ceiling

2026-06-118

api-integration

drewid74/ai_skills

Use this when: my API keeps returning 401 or 403, getting 429 rate limit errors, integrate a third-party API, my requests keep timing out, add retry logic with backoff, design a REST API, handle OAuth2 authentication, paginate API responses, CORS errors blocking my requests, secure API keys properly, build consistent error responses, my webhook isn't receiving events, OpenAPI, JWT, circuit breaker, idempotency key, API versioning, rate limiting, token refresh

2026-06-068

cicd-pipeline

drewid74/ai_skills

Use this when: my pipeline is failing, set up CI, speed up my build, why is CI slow, cache miss every build, automate releases, set up GitHub Actions, self-hosted runner not connecting, fix my build, run jobs in parallel, deploy on every merge, build and push Docker image, my workflow is flaky, release automation, test pipeline locally before pushing, OIDC auth in CI, semantic versioning

2026-06-068

code-reviewer

drewid74/ai_skills

Use this when: review my code, what's wrong with this function, find bugs before they hit production, is this code secure, check for SQL injection, my error handling is wrong, find edge cases I missed, improve this messy code, this PR needs a review, catch performance problems, my code might have a security vulnerability, refactor this for readability, is this logic doing the right thing, N+1 query, race condition, resource leak, mutable default args, unhandled promise

2026-06-068

data-engineering

drewid74/ai_skills

Use this when: build a data pipeline, my pipeline is not idempotent, clean messy data, convert CSV to Parquet, my data has duplicates, validate schema at ingestion, pipeline fails on re-run, process files larger than memory, schedule a recurring job, upstream schema changed and broke my pipeline, migrate data between systems, query Parquet without loading it, deduplicate records, batch vs stream processing, DuckDB for analytics, choose an orchestrator, slow pandas pipeline

2026-06-068

deploy-pipeline

drewid74/ai_skills

Use this when: get code onto my server, deploy to homelab, my deploy broke the service, set up automated deployment, zero-downtime deploy, how do I roll back, update Docker Compose services, health check failing after deploy, secrets in my deploy script, blue-green deployment, my service didn't come back up, automate deploys on git push, deploy without downtime, rsync to remote host, ZFS snapshot before deploy, Dockge stack deploy

2026-06-068

name

continue-from-accepted

description

Continue from accepted

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

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.