sustained-execution
Keeps long, multi-session, or compaction-spanning tasks on track using an external task ledger, checkpoints, and resumable state.
Skill by Adrian Paredez
Chalk Labs
Need help deploying sustained-execution?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/adrianparedez-sustained-execution
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: productivity
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/sustained-execution
What's inside
- Skill name
sustained-execution- Description
- Keeps long, multi-session, or compaction-spanning tasks on track using an external task ledger, checkpoints, and resumable state. Use when work spans many steps/hours, may hit context limits, or must survive a reset. Pairs with decomposing-tasks (the plan), context-budgeting (the window), and operating-autonomously (unattended runs).
- License
- Apache-2.0
- Compatibility
- Requires filesystem read+write for the ledger. Works on any runtime; benefits from but does not require automatic compaction.
- Metadata
- tier: 2 · layer: meta/control · author: Adrian Paredez · version: 1.0
- Instructions
- 85 lines in SKILL.md · view on GitHub
Files (7)
- SKILL.md
- README.md · benchmarks/
- resume.md · checklists/
- examples.md
- references.md
- checkpoint.md · templates/
- ledger.md · templates/
SKILL.md (first lines)
# Sustained execution Long tasks fail not from lack of capability but from **lost state**: the window fills, gets compacted, and the agent forgets where it was. The fix is an **external ledger** - durable memory that survives compaction and resets. The context window is scratch; the ledger is truth. ## Use this when - Work spans many steps, multiple sessions, or hours. - The run will likely hit the context hard ceiling (see `context-budgeting`). - The task must be **resumable** if interrupted. - **Skip** for short tasks that finish within one comfortable window. ## The ledger (single source of truth) Maintain one file (e.g. `notes/ledger.md`) holding: ``` GOAL + DONE criteria (verbatim, from decomposing-tasks) PLAN with step states ([x] done / [>] active / [ ] pending) DECISIONS log (what + why/constraint, append-only) CHECKPOINTS (digest pointers, newest last) ARTIFACTS (paths/URLs the run produced or depends on) NEXT ACTION (the single next concrete step, always current) ``` The **NEXT ACTION** line is the resume anchor: any fresh agent reads the ledger, finds NEXT ACTION, and continues without re-deriving anything. ## The loop ``` - [ ] INIT write GOAL+PLAN to the ledger (from decomposing-tasks) - [ ] before each step: read NEXT ACTION - [ ] do the step - [ ] after each step: update step state + DECISIONS + NEXT ACTION (write-through) - [ ] at context soft ceiling: offload detail, keep ledger pointers (context-budgeting) - [ ] at hard ceiling: write CHECKPOINT digest -> reinitialize -> reload goal+ledger - [ ] on resume/interrupt: read ledger -> NEXT ACTION -> continue - [ ] at end: confirm every DONE criterion; close the ledger ``` ## Principles 1. **Write-through, not write-back.** Update the ledger *immediately* after each step, not "later." A crash/compaction between step and write loses that step. 2. **The ledger is append-mostly.** DECISIONS and CHECKPOINTS append; PLAN step-states and NEXT ACTION mutate. Never rewrite history you might need. 3. **One current NEXT ACTION.** Always exactly one concrete next step recorded. If you can't write it, you don't know what to do next, re-plan. 4. **Checkpoints are resume points.** A checkpoint = a `context-budgeting` digest, saved, pointer added to the ledger. After a checkpoint, context holds only goal + ledger + active step. 5. **Verify before you record DONE.** A step isn't done because you ran it; it's done when its outcome passes its check (`verifying-reasoning` for critical ones). 6. **Idempotent resume.** Re-reading the ledger and re-doing the NEXT ACTION should be safe; design steps so a repeated step doesn't corrupt state (esp. one-way doors - record "done" only after the irreversible action confirmed). ## Resume protocol (the payoff) On any fresh start (new session, post-compaction, after interruption): ``` 1. Read the ledger. 2. Reload GOAL + DONE + PLAN states + latest CHECKPOINT + ARTIFACTS pointers. 3. Go to NEXT ACTION. Continue. (Do not re-explore solved sub-problems.) ``` ## Runtime adaptation - **Minimum:** filesystem read+write. The ledger is a plain Markdown file. - **If the runtime auto-compacts:** still keep the ledger, auto-summaries are lossy and not under your control; the ledger is. - **If a memory/state mechanism exists** (e.g. a memory dir, KV store): use it as the ledger store, same protocol. - **Multi-agent:** the ledger is the shared contract; each agent reads NEXT ACTION. ## Files - `references.md`, ledger schema, checkpoint cadence, crash-safety. - `examples.md`, a multi-session build walked end to end. - `templates/ledger.md`, the ledger to copy. - `templates/checkpoint.md`, checkpoint entry format. - `checklists/resume.md`, the resume protocol as a checklist. - `benchmarks/`, resumability and long-run completion method.
From AdrianParedez/capability-fabric, Apache-2.0.
More from AdrianParedez/capability-fabric
- 0.497score
auditing-skills
Adrian Paredez
Scores and improves an existing Agent Skill against a spec + quality rubric, then proposes or applies concrete refactors (split monoliths, sharpen…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
authoring-skills
Adrian Paredez
Generates a new, spec-compliant, portable Agent Skill from a capability description, frontmatter, focused body, and the supporting file set, following…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
composing-skills
Adrian Paredez
Routes a task to the smallest set of skills that covers it and sequences them without context bloat, instead of one monolithic instruction.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
context-budgeting
Adrian Paredez
Deliberately manages the context window to cut token cost and prevent quality decay on long runs.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
verifying-reasoning
Adrian Paredez
A fast, on-demand check that catches wrong answers before they ship, surface assumptions, seek disconfirming evidence, and validate the result against the…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
guarding-code-quality
Adrian Paredez
Keeps agent-generated code correct, safe, and consistent with the codebase, match existing conventions, verify behavior with a real run/test, and self-review…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
operating-autonomously
Adrian Paredez
Runs a goal unattended with a disciplined progress / stuck / done loop, bounded retry-with-variation, and clear stop and escalation criteria.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
progressive-research
Adrian Paredez
Finds external information without flooding context, a tiered search, triage, extract, corroborate, synthesize loop.
SkillCFreshness: C★ 1Claude Code +9 - 0.494score
decomposing-tasks
Adrian Paredez
Turns a goal into a verifiable, risk-ordered plan with explicit success criteria before any work starts.
SkillCFreshness: C★ 1Claude Code +9
Signals
- Stars
- 1
- Forks
- 0
- Open issues
- 0
- Last commit (this folder)
- Jun 5, 2026 (4 months ago)
- Last release
- Jun 5, 2026 (4 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- Version
- 1.0
- License
- Apache-2.0
- Website
- paredez.dev
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from Adrian Paredez
- 0.497score
auditing-skills
Adrian Paredez
Scores and improves an existing Agent Skill against a spec + quality rubric, then proposes or applies concrete refactors (split monoliths, sharpen…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
authoring-skills
Adrian Paredez
Generates a new, spec-compliant, portable Agent Skill from a capability description, frontmatter, focused body, and the supporting file set, following…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
composing-skills
Adrian Paredez
Routes a task to the smallest set of skills that covers it and sequences them without context bloat, instead of one monolithic instruction.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
context-budgeting
Adrian Paredez
Deliberately manages the context window to cut token cost and prevent quality decay on long runs.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
verifying-reasoning
Adrian Paredez
A fast, on-demand check that catches wrong answers before they ship, surface assumptions, seek disconfirming evidence, and validate the result against the…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
guarding-code-quality
Adrian Paredez
Keeps agent-generated code correct, safe, and consistent with the codebase, match existing conventions, verify behavior with a real run/test, and self-review…
SkillCFreshness: C★ 1Claude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get sustained-execution running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
sustained-execution skill. Agent Mart. https://agentmart.dev/skills/adrianparedez-sustained-execution (last updated 2026-10-24)