autospec-resume

autospec-resume (harness-neutral)

Detect an interrupted autospec run from durable run-state + heartbeats on a fresh start and auto-continue it. All decision logic routes through bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/resume-scan.sh" "$@".

Manage your own context — never exceed 60%. Delegate to subagents whenever your harness supports it.

<!-- autospec-block:startup-self-update SKILL_NAME=autospec-resume -->

Self-update mode

If the feature-request argument matches the regex ^\s*update\s*$ (case-insensitive, whitespace-padded), this skill enters self-update mode and does not run the normal pipeline:

1. Detect harness by checking which install path exists for this skill:
   • Claude Code: ~/.claude/skills/autospec-resume/SKILL.md
   • OpenCode: ~/.config/opencode/agent/autospec-resume.md
   • Codex CLI: ~/.codex/prompts/autospec-resume.md
2. Re-install the full autospec suite from main by piping the canonical installer:

   curl -fsSL https://raw.githubusercontent.com/berlinguyinca/autospec/main/bootstrap.sh | bash -s -- --skill all --harness all --update
   

   Run this one-liner once; it refreshes all autospec skills across all harnesses.
3. Show the diff between the prior installed file(s) and the freshly fetched copy.
4. Stop. Do not enter any pipeline phase. Print the upgrade summary and return to the user.

If no install path is detected, print Self-update: no installed copy of autospec-resume found; run install.sh first. and exit.

Harness detection

This skill is a thin dispatch wrapper — the relaunch happens through the durably-captured command, so it runs at the implementation tier. Detect your harness and pick the matching subagent tier:

• TIER_A (architecture / deep reasoning): not used here; this skill never plans.
• TIER_B (implementation work): the default — run the scan helper, print its one-line decision.

If your harness lacks a subagent primitive, run the helper inline and silently fall back to the in-context path rather than erroring. Never escalate a missing capability into a hard failure.

Stop mode

If the feature-request argument matches the regex ^\s*stop(\s+--\w+)*\s*$ (case-insensitive), defer to /autospec-stop: dispatch to bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/autospec-stop.sh" <args>, print its stdout, and do not enter any resume pipeline. Resume never halts a running monitor — that is /autospec-stop's job.

Invocation

/autospec-resume [--resume-partial] [--repo <owner/name>] [--dry-run]

Dispatches directly to bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/resume-scan.sh" "$@". All decision logic lives in the helper script.

Flag	Behaviour
(none)	Scan in-progress-by-bot and open auto-implement issues, apply the auto-resume pre-conditions, recover only exact stale-startup claims through Rust, and if all pass relaunch via the durably-captured command (clean-restart off origin/main).
--resume-partial	Additionally re-attach /tmp/wt-<branch> only when heartbeat.host == $(hostname); cross-host or missing host silently clean-restarts off origin/main.
--repo <owner/name>	Target a specific repo (used by the boot supervisor, which iterates the registry).
--dry-run	Print the recovery and relaunch actions that would run; invoke neither action; change nothing; exit 0.

Exit codes: 0 = nothing to resume / --dry-run / resumed (one-line reason printed); 1 = hard error (bad args, no gh).

Auto-resume contract (enforced by the helper)

• No second lock / idempotency. Resume relaunches the run; the relaunched monitor reads the existing GitHub CAS lock-comment through autospec claim state read (lowest-comment-id wins, loser self-cleans). Resume itself never writes a run-state comment and never adds any new lock. Two concurrent resumes (supervisor + human, or two hosts) converge to exactly one claim per issue at the existing CAS boundary.
• Pre-conditions (ALL required, else exit 0, no relaunch): (a) ≥1 crashed in-progress-by-bot issue or exact stale-startup candidate from open auto-implement; (b) no ~/.autospec/stop.flag; (c) no issue labeled paused-by-user; (d) not all issues closed.
• Stranded startup recovery. Scan both relevant labels, but accept auto-implement only when Rust run-state is the exact stale claimed / heartbeat-pending:none sentinel for the requested repository and issue. Delegate every ownership mutation to autospec claim state recover-stale-startup; relaunch only when Rust returns an exact matching recovered: true result.
• Crash-vs-live. Treat an issue as crashed (eligible) only when its age, computed from run-state server updated_at (never a local clock — mirrors autospec-watchdog.sh:386-405), satisfies step=claimed && age>=300 OR age>=10800. Otherwise assume a live/slow worker elsewhere and do not steal.
• Cross-host. --resume-partial re-attaches /tmp/wt-<branch> only when heartbeat.host == $(hostname); cross-host or missing host MUST clean-restart off origin/main (the crashed worktree is on a dead machine's local /tmp).
• Durable relaunch command. The command comes from the registry ~/.autospec/active-runs/<repo-slug>.json written by /autospec-run at monitor launch — never an attacker-supplied path. After a reboot (env cleared) the registry alone yields a runnable command.
• Attempt cap. Capped at AUTOSPEC_RESUME_MAX_ATTEMPTS (default 3) consecutive attempts without forward progress. At the cap, halt, print the cap-reached reason, and surface. The counter resets on forward progress (any issue merged).

Required capabilities & harness adapter

This workflow assumes a small set of capabilities. Map each one to your harness's actual tool. If a capability is missing, use the listed fallback.

Capability	Claude Code	OpenCode	Codex CLI	Fallback if missing
Run shell command	Bash	bash tool	shell / apply_patch	Ask user to run manually
Ask the user a question	AskUserQuestion	inline prompt	inline prompt	Ask in the response and wait for the next turn
Subagent model tier	Tier B: sonnet + medium thinking	Tier B: smaller-tier task + medium reasoning	Tier B: current spark/cost-optimized Codex + reasoning_effort=medium	Fall back UP on unavailability
<!-- autospec-block:harness-adapter-core -->

Model tier: Tier B (implementation work) — this skill is a thin dispatch wrapper.

Procedure

1. Run startup self-update block above.
2. Parse the user's invocation arguments.
3. Execute: bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/resume-scan.sh" "$@"
4. Print the helper's one-line decision to the user and exit.

If ${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/resume-scan.sh is not found, print:

autospec-resume: helper not found at ${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/resume-scan.sh.
Reinstall autospec-resume from the autospec repo, then retry.

and exit non-zero.

Hard rules

• Never write a run-state comment from this skill and never add a new lock. The relaunched monitor claims through the existing GitHub CAS lock only.
• Never invoke stale-startup recovery or the relaunch command under --dry-run.
• Never re-attach a cross-host or missing-host worktree under --resume-partial; clean-restart off origin/main.
• Never exceed AUTOSPEC_RESUME_MAX_ATTEMPTS consecutive auto-resume attempts; halt and surface at the cap.
• This skill is a thin wrapper. All behaviour is in the helper script.