CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

jbaruch/herdr-policy

Herdr team coordination policy for AI coding agents

74

Quality

93%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

bootstrap-recovery.mdskills/herdr-foreman/references/

Disappeared Codex Bootstrap Recovery

reconcile-bootstrap --record FILE records a transport failure and retires an owned shell after Codex exits during account/read TUI bootstrap. It sends no input, starts no replacement, changes no capabilities, and grants no correction or retry allowance. Run through the normal foreman entrypoint with the original --state; the command holds that owner's transaction lock.

The record requires dispatch, reason, authorization: {source, quote}, and evidence (absolute path to the saved visible pane text). Authorization cites the existing task decision. The foreman checks that the saved evidence is the original failure-only snapshot, taken before any shell interaction. Never replace it with a later snapshot containing other work. A missing name or report alone is not evidence that no turn began.

The saved text and first live pane observation must contain the complete frozen assignment beneath the literal Unsent draft: heading, immediately followed by Error: account/read failed during TUI bootstrap:. Terminal wrapping is accepted. The owner recomputes the prompt from its original role, COMMON, brief and tier, then checks its recorded prompt hash. Keep those frozen files and the saved pane text unchanged through recovery.

Normal assignment-scoped Codex apply now saves its original root shell and verified process before prompt input. Recovery requires that original pane and root shell, a sole-shell foreground, no registration on the pane, absent original agent, explicit null native-session observations before send, no started result, no other worker history, and no enrolled report. Recorded supervision observations of native work or a report also refuse recovery, even if later reads fail. An API failure is not absence. All these checks run again before cleanup. The first live pane must match the saved failure-only snapshot after ANSI removal. Its digest is persisted and must remain unchanged before every close; an appended shell command or output preserves the surface even when the old draft remains. If the original snapshot is unavailable or its provenance is uncertain, preserve the pane and reconcile the unknown work; a newly captured snapshot cannot establish the missing history. A changed pane, native session, report, foreign process or unknown transport preserves the surface.

For older dispatches, optional origin_evidence names an absolute path to an archived pre-exit herdr pane process-info JSON object (result and process_info envelopes are accepted). It must carry pane_id, shell_pid and one foreground_processes entry whose pid and argv equal the owner's original verified Codex process. The owner imports that receipt only after checking the current surface. A post-exit shell snapshot, pane title or agent name cannot substitute for this archive. Without original shell provenance, recovery refuses; preserve the pane and obtain actual launch evidence. Never add a shell PID by editing owner state.

The owner durably records outcome: not_sent before cleanup. Pending dispatches also receive the ordinary not_sent reconciliation, observed state absent and null native session. A legacy applied dispatch is eligible only with explicit started: false and no native/session/report proof. Its assignment, result, status and correction count remain historical facts; the separate bootstrap receipt records the transport correction. It does not make that applied row replayable or create retained context. Consult the owner receipt when recording the unavailable assignment in the task ledger; no report or approval is supplied.

Success emits the versioned receipt plus outcome: not_sent, retry_authorized: false, history_preserved: true and closure.closed: true. Exit 1 reports evidence, state or Herdr errors; exit 2 is CLI usage. If cleanup or supervision persistence fails, replay the identical record. Handle and acknowledge pending supervision observations through their existing owner gate before resolution. The transport receipt survives; replay retries only cleanup and resolution. A completed receipt replays without another close. Preserve unknown or sent work, and use normal planning and authorization gates for any later assignment.

skills

herdr-foreman

compose-briefs.sh

config.example.json

foreman-tier-check.py

foreman.sh

label-workspaces.sh

native-hook.sh

policy-paths.py

provision-worktree.sh

prune-report-caches.py

resolve-gates.sh

resolve-policy-paths.sh

review-package.sh

roster.sh

round-preflight.sh

SKILL.md

start-judge-worker.sh

state-schema.md

sweep-worktrees.sh

verify-authority.sh

wait-report.sh

README.md

tile.json