forge-recovery
Skill0 starsChanged 19 days agoMITView on GitHub
Use when an existing Forge run has actually crashed, stalled, or stranded work—for example stale lanes, dirty abandoned worktrees, unmerged work above the cap, no artifact progress, or product commits blocked behind REVISE. Do not load pre-emptively for a healthy run, a clean build, or plan review.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- is-bo/fullstack-forge-skilltheir own version, first seen Aug 28, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- is-bo/fullstack-forge-skill★ 2same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository5 other files it gives its agents.
- forge-buildSkill
- forge-orchestrateSkill
- forge-reviewSkill
- forgeSkill
- forge-with-docsSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 12
Load only when recovery is actually needed: If the run is progressing, this skill is noise — use /forge.
- Very specific: line 69
Commit check is: changed paths match the admitted set · no overlap/safety breach · executor proof exited 0.
- Distinctive: line 91
Additive integration only: No secrets in diffs, logs, commits or reports.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
FORGE-RECOVERY — Jump-Start a Wrecked Run, Then Resume the Plan
/forge assumes a clean start. This assumes wreckage: dirty trees, stranded
commits, unreceipted claims, a blocked machine, an orchestrator that is busy but not
progressing, and a control plane that permitted it.
Load only when recovery is actually needed. If the run is progressing, this skill
is noise — use /forge.
Recovery is for a plan that existed and went wrong. Healing it and getting back on track is the job; closing an open review round is part of Act 4 only when one is genuinely open.
The acts
Act 0 is not a step — it is armed for the whole run and fires inside any other act.
| Act | Does | Objective metric (what the watchdog measures) |
| 0 WATCH | Outcome-based stuck detection + role handoff | (armed throughout) |
| 1 HALT & PRESERVE | Freeze dispatch; adjudicate live-vs-hung by the artifact clock; patch-preserve every dirty tree; name every detached tip | preservation receipts written |
| 2 RECONSTRUCT | Rebuild ground truth from artifacts; void every unreceipted claim; SHA-dedupe the commit union | State Ledger sections completed |
| 3 REPAIR | Fix the control plane that permitted the failure; add the refusal test; append the lesson | guard + watchdog self-tests passing |
| 4 RE-LEGALIZE | Close any genuinely open review round; write the salvage ledger; get it adversarially approved | critique rounds closed |
| 5 DRAIN | Land salvage units under the guard, cheapest proof first | units landed + gate state |
| 6 REJOIN | Hand back to forge at its true act with a self-contained resume prompt | resume prompt written |
Full mechanics, gates and per-act procedure:
references/acts-detail.md
Act 0 — the watchdog, in brief
./recovery-watchdog.sh measures outcomes, not activity. sol-liveness.sh answers
"is this process alive?"; on 2026-07-29 the answer was yes for twelve hours while
nothing landed.
WD="${FORGE_SKILLS:-$HOME/.claude/skills}/forge-recovery/recovery-watchdog.sh" # FORGE_SKILLS: plugin installs
$WD init --run "$RUN" --unit R1 --owner <side> --peer <side> --objective "units landed"
$WD attempt --run "$RUN" --unit R1 --result pass|fail --detail "<verbatim failure>"
$WD progress --run "$RUN" --unit R1 --evidence "<what objectively moved>"
$WD check --run "$RUN" --unit R1 # 0 OK | 10 handoff | 20 escalate | 30 unknown| Trip | Action |
REPEAT — 2 consecutive failed attempts under one owner | hand the stuck role to an eligible peer |
STALL — no objective progress for 20 min | hand over |
BLOCKED — a precondition refusing for an hour | hand over, block declared |
PINGPONG — every eligible owner has had a real tenure | escalate: one named ask plus the default action you will take anyway |
Handoff rules: one owner per unit ever · only the stuck role transfers · the taker
may change approach but the proof command is immutable · the active owner overlay
still binds (under Codex-only, product execution transfers to another Codex executor).
Template: HANDOFF-PACKET.md.
Cadence
Recovery honours the active execution profile exactly as
/forge does.
sprint— one smallest relevant proof per unit; no unit review, no complete-diff review, no orchestrator proof rerun, no fresh-database rebuild, no full unit gate. Commit check is: changed paths match the admitted set · no overlap/safety breach · executor proof exited 0. A red targeted proof is fixed and re-run — it does not by itself selectassured. Overlay/default executor is applied automatically, one line recorded, no routing debate and no Gate-2 wait.assured— per-unit cross-model review with receipts, complete-diff review, the orchestrator's own proof rerun, the integrated gate, argued routing and Gate 2.
Recovery mode relaxes only the clean-tree precondition — never receipts required at the active profile, roles, caps, leases, machine admission, or the rule that the executor never commits.
Safety invariants — binding at every profile
- Owner instructions win, always.
- Preserve before anything. No
reset,clean, destructiverestore, branch deletion, force-push, or rebase of landed commits without explicit owner approval — and never before Act 1's receipts exist. - No lane launches without
dispatch-guard.shprintingADMIT:. But a refusing guard must route to safe work and, pastENV_MAX, hand over — never idle. - Three lanes total; one heavy lease; disjoint declared file sets; one writer per worktree.
- The executor never commits or merges. The orchestrator does.
- Additive integration only. No secrets in diffs, logs, commits or reports.
- Claims are not facts — no receipt, no credit.
Reference
| Need | File |
| Per-act mechanics and gates | references/acts-detail.md |
| What actually happened, with numbers | references/incident-2026-07-28.md · ../forge/references/incidents.md |
| Loadable orchestration rules | PLAYBOOK.md |
| Handoff template | HANDOFF-PACKET.md |
| Routing | ../forge/ROUTING-POLICY.md |
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

