Skip to content
Agentmart
Submit a tool

deep-finish

Use when finishing an authorized deep-work session with keep, PR, merge or discard and a durable receipt.

Skill by Sungmin-Cho

IndexedMIT★ 10
Write a review

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/sungmin-cho-deep-finish

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.

We use your email only for this newsletter. Privacy

About

Use when finishing an authorized deep-work session with keep, PR, merge or discard and a durable receipt.

No 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/deep-finish

What's inside

Skill name
deep-finish
Description
Use when finishing an authorized deep-work session with keep, PR, merge or discard and a durable receipt.
License
MIT
Instructions
21 lines in SKILL.md · view on GitHub

Files (5)

  • SKILL.md
  • completion-options.md · references/
  • handoff-emit.md · references/
  • phase5-markers.md · references/
  • session-quality-score.md · references/

SKILL.md (first lines)

# Finish with durable evidence

Resolve the plugin root to a literal absolute path. Before reading or running any plugin file, resolve its real path and require containment in that root. All `${CLAUDE_PLUGIN_ROOT}` paths below mean that verified root.

Resolve `--session=ID` from `$ARGUMENTS`; otherwise run `node "${CLAUDE_PLUGIN_ROOT}/scripts/deep-work-runtime.js" session context`. Read the returned state path and its `work_dir`; never guess a work directory or change the active pointer to another session. Validate with `session authority validate --state "$STATE_FILE"`. Use the runtime's phase-aware result: a legitimate pre-plan session does not require a plan that has not been created.

Read `${CLAUDE_PLUGIN_ROOT}/skills/shared/references/runtime-execution-spine.md`. The runtime owns phase, approval, write, review and receipt mutations. Run `node "${CLAUDE_PLUGIN_ROOT}/scripts/deep-work-runtime.js" --help` for exact current route arguments; consume returned IDs/digests rather than inventing them.

Resolve outcome from explicit arguments or existing user authorization: merge, pr, keep or discard. Ask for an outcome only when it is genuinely unspecified and cannot be inferred. Discard remains destructive and must be actually authorized. Internal continuation preferences do not grant push/PR/merge/publication authority.

Use the runtime's authenticated completion/metrics projection. Read `${CLAUDE_PLUGIN_ROOT}/skills/deep-finish/references/session-quality-score.md` for interpretation. Create any required payload/title/body as owned temps; never write a completed session receipt or finalized history record by hand.

Invoke the exact `session finish keep`, `session finish publish-pr`, `session finish merge` or `session finish discard` route from runtime help. It validates readiness before effects and records their intent/results. A post-effect verification or receipt failure resumes from the stored effect identity; never issue a duplicate PR, merge or discard to recover a missing receipt.

The runtime publishes the stable session M3 envelope through the contained canonical wrapper before finalizing state, registry and pointer. New slice/session payloads use registry version 1.1; the top envelope format remains 1.0. Do not invoke the wrapper with a handcrafted completion payload to bypass the runtime.

Verify the terminal receipt and actual requested repository/external outcome. Preserve the active worktree and evidence unless cleanup was requested. Optional memory/wiki ingestion follows explicit applicable authorization and does not substitute for goal completion.

Final response: outcome, artifact/PR/release links, relevant verification and concrete limitations. Do not claim success while publication or verification remains pending.

From Sungmin-Cho/deep-work, MIT.

More from Sungmin-Cho/deep-work

Signals

Stars
10
Forks
2
Open issues
1
Last commit (this folder)
Sep 7, 2026 (1 month ago)
Last release
Sep 20, 2026 (1 month ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from Sungmin-Cho

Cite this page

deep-finish skill. Agent Mart. https://agentmart.dev/skills/sungmin-cho-deep-finish (last updated 2026-10-24)