Agent Triforge
Install
Sessionlead workflow

at-resume

Restore STATE.md, re-prove the baseline when HEAD moved, summarize, continue the recorded phase.

Use when starting a session on existing work: restores ops/STATE.md, re-proves the baseline and continues the phase.

SKILL.md on GitHubReference files on GitHub

Run it

A workflow runs in the agent that leads. Type it in that agent's session.

Claude Code

/at-resume
On this page

Resume

Goal: the sprint continues from where the last session stopped, with the saved state proven current rather than assumed. Done when the user has the summary (sprint goal, current phase, tasks completed vs remaining, blockers or pending decisions, recommended next action), has said whether to continue automatically or review the state first, and work has resumed at the recorded phase. Safe failure: when ops/STATE.md is missing, say so and stop; a stale baseline is re-proven before anything else; a merged lease is never redone; when the saved state and the tree disagree, ask before acting.

Invoked with no arguments. The user's instructions outrank the saved plan: a user who says what to do next overrides "Next actions".

This is the session-continuity skill's resume protocol (full text: $ROOT/skills/session-continuity/SKILL.md). The helper and the sibling skills are reached through ROOT=$(bash "$SKILL_DIR/scripts/locate-triforge.sh") || exit $?; source "$ROOT/scripts/invoke-external.sh"; the locator fails closed naming at-setup. $SKILL_DIR is the directory this SKILL.md was loaded from — the harness shows that path when it loads the skill (the plugin install under Claude Code, the skill's path under Codex) — and every path in this skill is relative to it, never to the project; never run a project's own scripts/locate-triforge.sh.

Step 1: Load state

Read, in order: ops/STATE.md (where you left off — phase, progress, next actions; a Type: checkpoint marker means a pause, not a wrap), ops/TASKS.md (task status), ops/MEMORY.md (decisions and gotchas from previous sessions), ops/CHANGELOG.md (what was already done), ops/CONTRACTS.md (current interface definitions).

Step 2: Assess the current state

  • Baseline: when git rev-parse HEAD differs from state_head, or verification_baseline.commit differs from HEAD, or the baseline is missing, re-run verification_command and record the fresh result before continuing. A baseline from another commit proves nothing about this one.
  • Uncommitted changes (git status); the phase active at pause; remaining tasks by status (active, in progress, blocked, review); review files present (ops/REVIEW_ANTIGRAVITY.md, ops/REVIEW_CODEX.md, ops/TEST_RESULTS.md).
  • Leases: when ops/leases.toml exists, lease_heartbeat_check reclaims orphans (rc 44 means a git-integrity change the lead did not make — inspect it, then lease_rebaseline); open leases are requeued or finished, building ones waited on with lease_wait as the harness reference says (references/claude.md, references/codex.md); merged leases are already on the integration branch and are never redone. See references/lease-resume.md.

Step 3: Report to the user

Summarise: sprint goal, current phase, tasks completed vs remaining, blockers or pending decisions, recommended next action. Then ask whether to continue automatically or review the state first, and wait.

Step 4: Resume execution

Pick up from the phase recorded in ops/STATE.md: mid-Phase 2 → the remaining build tasks under the wave protocol (wave-orchestration, driven by at-build); mid-Phase 3–4 → check whether the reviews are complete and process them (at-review); mid-Phase 5 → check the test results and fix failures (at-test); Phase 6 needed → at-wrap. Those workflows are user-only: tell the user to run the one named (/at-build, or $agent-triforge:at-build in a Codex prompt, and so on); if you must continue yourself, read its skills/<name>/SKILL.md under the plugin root that bash "$SKILL_DIR/scripts/locate-triforge.sh" prints, and follow it. Sub-agents a resumed phase needs are spawned through the lead's harness, model and effort pinned on every spawn — see references/claude.md and references/codex.md.

Output

No artifact of its own. The summary to the user (sprint goal, current phase, tasks completed vs remaining, blockers or pending decisions, recommended next action), the fresh baseline recorded in ops/STATE.md when HEAD had moved, and then the resumed phase's own outputs.