Loop Execution

HOTL Loop Execution

Overview

Execute an existing HOTL workflow file autonomously. Canonical workflows live at docs/plans/YYYY-MM-DD-<slug>-workflow.md, with legacy root hotl-workflow-<slug>.md still accepted during migration. Loop on steps with success criteria. Auto-approve low-risk gates. Always pause for high-risk gates.

governed-execution is the preferred new router for host-native or fallback runs. This skill remains the canonical autonomous state machine and a supported explicit compatibility profile. When invoked through the router, use the selected driver for lifecycle calls and require a sufficient evidence receipt before completion.

Announce: "Starting HOTL loop execution. Looking for workflow file..."

Workflow File Resolution

Resolve which workflow file to execute:

1. If the user specified a filename → use that file
2. Else, glob for canonical workflows in docs/plans/*-workflow.md:
   • One match → use it automatically
   • Multiple matches → if one is clearly the newest revision for the same semantic slug, prefer it; otherwise list them and ask the user to pick
3. If no canonical matches, glob for legacy hotl-workflow*.md in project root:
   • One match → use it automatically
   • Multiple matches → list them and ask the user to pick
4. No matches → see "What to Do If No Workflow Found" below

Interrupted Run Detection

After resolving the workflow file, locate interrupted runs before Branch/Worktree Preflight. Use scripts/hotl-locate-run.sh --workflow <workflow-file> when available. This helper scans the current checkout, linked git worktrees, and HOTL's default .hotl-worktrees/<repo>/ directory, so it can find state created inside an isolated execution worktree even when the new session starts from the authoring checkout.

If the helper is unavailable, manually scan these locations for .hotl/state/*.json and match workflow_path or source_workflow_path against the resolved workflow:

1. Current checkout
2. git worktree list --porcelain roots for the repo
3. $(dirname <repo-root>)/.hotl-worktrees/$(basename <repo-root>)/*

• One interrupted run found → ask: "Found an interrupted run (step N/M). Resume from step N, or start fresh?"
• Multiple interrupted runs found → list all with run_id, step progress, branch, age. Ask which to resume or start fresh. Never silently choose.
• No interrupted runs → proceed normally (new run)

If the user chooses resume, do not run Branch/Worktree Preflight again. Load the matched sidecar, change into its recorded execution_root, and follow skills/resuming/SKILL.md with the original run_id.

Branch/Worktree Preflight

After resolving the workflow file, run this preflight before executing any steps:

1. Is this a git repo with at least one commit?
   - No  → log "Skipping branch setup (no git history)" → proceed to step execution
   - Yes → continue

2. Check for uncommitted changes
   - First, exclude HOTL-owned transient artifacts from the dirty check:
     • docs/plans/*-workflow.md (canonical workflow files)
     • hotl-workflow-*.md (legacy workflow files)
     • docs/designs/*.md (canonical design docs from brainstorming)
     • docs/plans/*-design.md, docs/plans/*-plan.md (legacy design docs from brainstorming)
     • .hotl/ (runtime state, reports, cache)
   - If only HOTL artifacts are dirty → treat as clean, continue
   - If non-HOTL dirty files exist:
     • If dirty_worktree: allow in workflow frontmatter → proceed without prompting
     • Otherwise → HARD-FAIL. Tell the user which non-HOTL files are dirty. Offer choices:
       a. Clean up manually, then re-run
       b. Stash manually, then re-run
       c. Explicitly approve HOTL to stash and continue
   - Clean → continue

3. Determine branch name
   - If branch: field exists in workflow frontmatter → use it
   - Otherwise → derive `hotl/<slug>` from the workflow filename
     • Canonical: strip `YYYY-MM-DD-` prefix and `-workflow.md` suffix from `docs/plans/YYYY-MM-DD-<slug>-workflow.md`
     • Legacy: strip `hotl-workflow-` prefix and `.md` suffix from `hotl-workflow-<slug>.md`

4. Capture authoring origin
   - Record the current branch name (if any) and current `HEAD` commit as the workflow's authoring origin
   - If the current branch is neither `main` nor `master`, and the workflow frontmatter does not already set `branch:` or `worktree:`, PAUSE and ask:
     a. Continue on the current branch in this checkout
        → set `branch: <current-branch>` and `worktree: false`
     b. Use HOTL's isolated execution branch/worktree (recommended)
        → leave `worktree: true` and let HOTL derive `hotl/<slug>` unless the user wants a custom branch name
     c. Use a custom execution branch
        → set `branch: <user-branch>` and keep worktree isolation unless the user explicitly opts out
   - Explain clearly: the authoring checkout and the execution checkout can differ. HOTL can execute in a separate worktree while leaving the current checkout untouched

5. Determine isolation mode
   - If `worktree: host` in frontmatter → stay on the current checkout's current feature branch exactly as provided by the host tool; reject `main` and `master`
   - If `worktree: false` in frontmatter → stay in the current checkout and use a dedicated branch there
   - If the current checkout is already a named linked git worktree, and the workflow frontmatter does not set `branch:` or `worktree:`, use host mode automatically to avoid stacking another worktree
   - Otherwise → use an isolated git worktree by default

6. Check if the target branch/worktree already exists locally
   - Current helper behavior: existing branch/worktree collisions are a hard stop, not an interactive reuse/recreate flow
   - If the helper reports an existing branch/worktree conflict, stop and ask the user whether to reuse manually, delete+recreate manually, or abort
   - Does not exist → create (no prompt)

7. Resolve the execution root with `scripts/hotl-prepare-execution-root.sh <workflow-file> --executor-mode <mode>`
   - The helper returns JSON with: `branch`, `repo_root`, `execution_root`, `workflow_path`, `source_workflow_path`, `source_branch`, `source_head`, `worktree_path`
   - By default it creates a linked git worktree for the branch, copies the current workflow into that worktree, and returns that worktree as `execution_root`
   - If `worktree: false` in frontmatter → create/switch to the dedicated branch in the current checkout and return the repo root as `execution_root`
   - If `worktree: host` in frontmatter → keep the current branch and return the current checkout as `execution_root`; if `branch:` is set, it must match the current branch
   - If `branch:` matches the currently checked-out branch while worktree isolation is still enabled, the helper must STOP with a clear message telling the user to set `worktree: false` or `worktree: host` for same-branch continuity
8. Change into `execution_root`
   - Every later git command, runtime call, Codex helper call, and review command for this run MUST execute from that directory

Rules:

• No auto-stash. Hidden state mutation weakens governance.
• Existing branch/worktree collisions do not currently auto-prompt; treat them as a stop-and-ask condition until interactive reuse/recreate is implemented.
• Non-git repos skip entirely — HOTL works without git ceremony.
• Run HOTL structural lint (scripts/document-lint.sh) automatically on the workflow file before any git mutation or step execution. If lint fails, STOP and show all errors. If lint passes, continue silently.

HOTL Execution State Machine

This is the canonical HOTL execution state machine. Other execution modes (e.g., subagent-execution) reference this spec and define only their differences.

1. Resolve workflow file (see above)
2. Parse frontmatter: intent, risk_level, auto_approve, branch, worktree
3. Run Branch/Worktree Preflight (see above)
4. Capture preflight metadata and change into `execution_root`
5. Initialize and claim the run via runtime:
   - Run: `hotl-rt init <workflow-file> --require-owner --executor-mode loop --repo-root <repo-root> --execution-root <execution-root> --source-workflow-path <source-workflow-path> --source-branch <source-branch|null> --source-head <source-head|null> --worktree-path <worktree-path|null> --branch <branch>`
   - This parses the workflow, creates .hotl/state/<run-id>.json with all steps, and initializes .hotl/reports/<run-id>.md
   - Capture the run_id from stdout
   - Immediately run `hotl-rt owner claim --owner <stable-controller-id> --lease-seconds <bounded-lease> --run-id <run-id>`. Parse the one-time `token` from its JSON response, retain it only in the controller, export it as `HOTL_OWNER_TOKEN`, and clear any temporary response variable. Never write the raw token to chat, reports, prompts, or committed files.
   - Run `hotl-rt owner heartbeat --lease-seconds <bounded-lease> --run-id <run-id>` before and after long actions and at safe transition boundaries. Use explicit `owner handoff`, `owner release`, or reviewed `owner takeover`; timestamp age alone never grants ownership.
   - For every later mutating `hotl-rt step`, `hotl-rt gate`, `hotl-rt action`, `hotl-rt budget`, `hotl-rt finalize`, `hotl-rt finish`, and owner call, keep `HOTL_OWNER_TOKEN` in the controller environment. Pass `--run-id <run-id>` or set `HOTL_RUN_ID=<run-id>` for all runtime and helper calls.
   - Never let the runtime silently choose "the newest run" when multiple runs exist
   - Only after init and `owner claim` succeed should chat output or native plan/progress UI show anything

6. For each step in order:

   a. Start step via runtime:
      - Run: `hotl-rt step N start --run-id <run-id>`
      - This persists step start (status, timestamp, attempts) to state and report
      - Only after the runtime call succeeds should chat show "→ Step N"

   b. Announce: "→ Step N: [name]"

   c. Execute the action (agent implements the work)
      - Heartbeat immediately before starting and after returning from any action that may run for a meaningful fraction of the lease.
      - Host goals, automations, background sessions, handoffs, and hooks are scheduling and liveness only. The controller must still persist every transition through HOTL. Preview and experimental host continuation features are opt-in.
      - For `external_write`, `production_change`, or `secret_access`, first run `hotl-rt action request <kind> --target <bounded-target> --idempotency-key <stable-key> --run-id <run-id>`, obtain the required human decision with `action decide`, then persist intent with `action begin <action-id> --idempotency-key <stable-key> --run-id <run-id>` before performing the effect.
      - After observing the effect, run `action complete <action-id> <succeeded|failed|cancelled> --evidence-ref <reference> --run-id <run-id>`. If execution was interrupted or the outcome is uncertain, inspect the target and use `action reconcile`; never replay an `in_progress` or `uncertain` effect blindly.

   d. Verify via runtime:
      - Run: `hotl-rt step N verify --run-id <run-id>`
      - The runtime runs the verify command, captures stdout/stderr, and atomically transitions the step to done or failed
      - If the verify type is unsupported, the runtime blocks the step with a clear reason
      - For type: browser — `hotl-rt` downgrades to a paused human-review request when browser tooling is unavailable inside the runtime. If the controller has platform browser tooling, perform the browser check before approving or rejecting that pause.
      - For type: human-review — the runtime returns a `human review required: ...` block reason and sets the run status to `paused`; ALWAYS pause for human (never auto-approve)
      - For type: artifact — runtime checks path exists and evaluates assert; for `matches-glob`, `path` must be the directory and `value` must be a filename glob only, so `src/*` is invalid and should be authored as `path: src`
      - Runtime transitions enforce step order, per-step `max_iterations`, aggregate attempt and elapsed-time limits, and configured budget pauses. Run `hotl-rt budget check --run-id <run-id>` when external agent or cost telemetry was recorded or before launching another costly/long action.

   e. If verify fails (runtime returns non-zero):

      e0. If the runtime output says `blocked: human review required: ...`
         → PAUSE immediately. Do not start later steps or finalize the run.
         → Show the review prompt to the human and ask: "Continue? (yes/no/show-details)"
         → If the human says yes/approve/continue:
             Run: `hotl-rt gate N approved --mode human --run-id <run-id>`
             Then continue to the next step
         → If the human says no/reject:
             Run: `hotl-rt gate N rejected --mode human --run-id <run-id>`
             STOP and surface the report path
         → If the human asks for details:
             Show the relevant test/report context, then wait again
         → Never treat the chat reply alone as persisted approval; the approval is only real after the `hotl-rt gate ...` call succeeds

      f. If loop: false
         → STOP, report to human
         → Run: `hotl-rt step N block --reason "verify failed" --run-id <run-id>` if not already marked failed by verify
         → Show last verify output. Wait for human guidance.

      g. If loop: until [condition]
         → if iterations < max_iterations:
             Run: `hotl-rt step N retry --run-id <run-id>` then `hotl-rt step N start --run-id <run-id>`
             log "↻ Retrying ([n]/[max])...", retry the action
         → if iterations = max_iterations: STOP
             Report: "Step N reached max iterations ([max]). [condition] not met."
             Show last verify output. Wait for human guidance.

      h. On step completion (verify passed):
      - The runtime has already persisted the done status
      - Update the workflow checkbox to [x]
      - Only after the runtime confirms success should chat show "✓ Step N"

   i. If gate: human
      → if auto_approve: true AND risk_level != high:
          Run: `hotl-rt gate N approved --mode auto --run-id <run-id>`
          log "⚡ Auto-approved: Step N gate (risk: [risk_level])"
          continue
      → else:
          PAUSE. Show summary of what was done in this step.
          Ask: "Gate reached at Step N. Continue? (yes/no/show-details)"
          Wait for human response.
          Run: `hotl-rt gate N approved --mode human --run-id <run-id>` or `hotl-rt gate N rejected --mode human --run-id <run-id>`

   j. If gate: auto
      → Run: `hotl-rt gate N approved --mode auto --run-id <run-id>`
      → always continue, log "⚡ Auto-approved: Step N gate"

7. All steps complete:
   → Run review checkpoint (see Review Checkpoints below)
   → Invoke hotl:verification-before-completion skill
   → For Codex final summaries, run: `scripts/finalize-codex-summary.sh <run-id>` (or set `HOTL_RUN_ID`)
   → For Claude Code/Cline, run: `hotl-rt finalize --json --run-id <run-id>`, write the payload to a temp file, then render it with: `scripts/render-execution-summary.sh --platform <claude|cline> <summary-json-file>`
   → Successful finalize sets status to `ready_to_finish`, not `completed`. It proves step/gate/effect/budget readiness but does not invent a branch disposition.
   → Do not freehand the final summary when the renderer is available
   → Invoke `hotl:finishing-a-development-branch` with the same `run_id` and retain `HOTL_OWNER_TOKEN`. Even non-git runs require an explicit `kept` disposition before completion.
   → Never silently merge, publish, keep, or discard the execution branch/worktree
   → After `hotl-rt finish` moves a successful run from `ready_to_finish` to `completed`, run `owner release`, then require `hotl-rt receipt <run-id> --json` to report `sufficiency.sufficient: true`.
   → Only then show the rendered summary as visible chat output in the final response; do not paraphrase it away or claim completion earlier.

Execution State Persistence

All state persistence is handled by the hotl-rt shared runtime (runtime/hotl-rt). Agents do not manage state files directly.

The runtime owns:

• .hotl/state/<run-id>.json — authoritative machine state (created by hotl-rt init, updated by owner/step/gate/action/budget/finalize/finish transitions)
• .hotl/reports/<run-id>.md — durable Markdown report (initialized at init, updated incrementally, finalized at finalize)

The sidecar also records the execution location for the run: repo_root, execution_root, workflow_path, source_workflow_path, worktree_path, and executor_mode.

Run ID format: <slug>-<YYYYMMDDTHHMMSSZ>-<12-hex-nonce> (e.g., add-auth-20260320T212315Z-a1b2c3d4e5f6). The entropy suffix prevents collisions across independent execution roots; legacy timestamp-only IDs remain readable.

Workflow checkboxes (- [x]) are a human-visible mirror updated by the agent on step completion. The sidecar is the source of truth.

Operational rule: hotl-rt calls happen before the corresponding chat log or Codex native plan/progress update. Native progress UI is never a substitute for the runtime-managed artifacts.

Concurrency rule: after hotl-rt init returns a run id, claim the renewable controller and pin every later runtime/helper call to that run id with HOTL_OWNER_TOKEN. Do not rely on "latest file in .hotl/state" when more than one run exists, and do not use file age as takeover authority.

See skills/resuming/SKILL.md for the full sidecar schema, stale run detection, and verify-first resume flow.

Path Resolution

To find hotl-rt and HOTL scripts (document-lint.sh, hotl-locate-run.sh, render-execution-summary.sh, etc.), resolve in this order:

1. Session context (Claude Code): the session-start hook injects the plugin base path — use it to construct full paths like bash <plugin-path>/runtime/hotl-rt init <workflow-file>
2. Codex native-skills install: resolve from ~/.codex/hotl/runtime/hotl-rt and ~/.codex/hotl/scripts/
3. Codex plugin install: resolve from ~/.codex/plugins/hotl-source/runtime/hotl-rt and ~/.codex/plugins/hotl-source/scripts/
4. Codex plugin cache fallback: resolve from ~/.codex/plugins/cache/codex-plugins/hotl/*/runtime/hotl-rt and ~/.codex/plugins/cache/codex-plugins/hotl/*/scripts/
5. Cline: resolve from ~/Documents/Cline/Scripts/hotl-rt and ~/Documents/Cline/Scripts/
6. Working in the hotl-plugin repo itself: use ./runtime/hotl-rt and ./scripts/

The same resolution applies to all HOTL scripts under scripts/.

Dependency: jq. hotl-rt requires jq for JSON state management. If hotl-rt fails with a jq not found error:

1. Show the user the platform-specific install command from the error output
2. Fall back to inline execution — run steps with per-step chat logs and manual checkbox updates, but without state persistence (.hotl/state/), durable reports (.hotl/reports/), or deterministic summary rendering
3. Note in the final output that the run was not state-managed due to missing jq

Execution Report

Execution report output must conform to docs/contracts/execution-report-output.md. The contract defines the durable report format (metadata, summary table, event log), execution status vocabulary, final summary semantics, platform rendering tables, and the deterministic renderer reference.

The hotl-rt runtime writes the durable report to .hotl/reports/<run-id>.md incrementally. The report survives app rendering quirks and provides a reliable post-run artifact for debugging, trust, and resume.

Reference report_path in user-facing pause, blocked, resume, and completion responses so the durable report is always discoverable.

When a verify: human-review step pauses, the response must include the report_path and make it clear that the run is paused pending approval, not failed.

If the workflow sets report_detail: full, successful verify output must also be included in the durable report, not only failures.

Review Checkpoints

Record git rev-parse HEAD as the review base at run start and after each review.

At Final Completion

After all steps have passed verification, before hotl-rt finalize:

1. A final review is required unless the most recent review already covers all current changes and no code changed afterward
2. Invoke requesting-code-review with review type: final
   • Review base: branch point or last review base, whichever is more recent
3. When findings return, invoke receiving-code-review
   • Follow Verify → Evaluate → Respond → Implement
4. Resolve all BLOCK findings before finalizing
5. If fixes after the last review changed scope, constraints, or risk_level, request a scoped follow-up review before completing

At Intermediate Gates (Conditional)

At intermediate gate: human steps, request review only when:

• risk_level is high
• The gate covers cross-module or feature-scale changes
• Multiple steps have completed since the last review

When triggered, scope the review to steps completed since the last review checkpoint, not the entire run. Use review type: checkpoint.

Do not request review at every gate by default.

Review Ordering

Review happens:

1. After step verification is complete
2. Before verification-before-completion
3. Before hotl-rt finalize / any "done" claim

Safety Rules

• risk_level: high in frontmatter always forces human approval at gate: human steps, even if auto_approve: true
• Never skip a gate: human on steps with security-sensitive keywords (auth, encrypt, secret, key, password, token, permission, role, billing)
• On STOP: always show the failing verify output so human can diagnose

Live Execution Behavior

Report format, status vocabulary, final summary semantics, and platform rendering tables for final artifacts are defined in docs/contracts/execution-report-output.md. This section covers runtime behavior that is executor-owned: live step visibility, progress updates, and verbose mode.

Final Summary (mandatory)

Every execution run MUST end with a visible final summary in chat. A prose recap alone is not compliant.

For Codex final summaries:

• MUST use scripts/finalize-codex-summary.sh when available
• MUST include the rendered compact list in the final response as visible chat text
• MUST keep the rendered step lines intact; do not paraphrase, omit, or replace them with narrative
• MAY add a short prose recap after the rendered summary, but not instead of it

If the Codex helper is unavailable, fall back to hotl-rt finalize --json plus scripts/render-execution-summary.sh --platform codex ..., then emit that renderer output directly.

Platform Rendering

Final artifacts must follow docs/contracts/execution-report-output.md:

Status vocabulary for final summaries includes ✓ Done, ⚡ Auto-approved, ✓ Approved, ✗ Failed, and ✗ Blocked.

Iterations means attempt count only. For tables, the Iterations column is a number only or - for gates. Never put test counts in Iterations; test counts belong in Status.

In the Codex compact list, keep the step name first and inline status detail after -. Include the status word on every line and always include iteration count details such as Done (1 attempt), Done (3 attempts), or Approved (1 attempt).

Platform Live Step Visibility

Live Step Visibility (mandatory)

Every execution run MUST provide live step visibility — the user must see which step is currently executing and which are done. This is not optional on any platform.

Codex Native Progress (mandatory with fallback)

When running in the Codex app, the executor MUST use the native plan/progress UI as the primary live step visibility surface:

• MUST initialize the native progress card immediately after run setup (step 4 of the state machine)
• MUST update it on every step transition — exactly one step in_progress at a time
• MUST keep the native card high-level: major workflow steps only, not retries or verify substeps
• If the native progress tool is unavailable or errors, MUST immediately switch to per-step chat logs for the remainder of the run. Do not silently drop visibility.
• Native progress never replaces the final chat summary, durable report, or sidecar state
• If Codex needs an explicit readout of the active workflow step, use scripts/show-codex-current-step.sh to print the current step number, name, status, and attempts from the active run without mutating state

On platforms without native progress (Claude Code, Cline), the executor MUST use per-step chat logs for live visibility.

Per-Step Log (default, always shown)

After each step, log one line:

✓ Step 1: Write failing tests
✓ Step 2: Implement auth logic (3 attempts)
⚡ Step 3: Security review gate (auto-approved)
✓ Step 4: Update docs

Verbose Progress View (opt-in)

When verbose mode is enabled, print a compact step list at each step transition (before starting a step, after a step completes/fails/auto-approves):

  ✓ Step 1: Write failing tests
  ✓ Step 2: Implement feature
  → Step 3: Run full test suite (attempt 1/3)
  · Step 4: Update docs
  · Step 5: Human review

Symbols:

• ✓ — completed
• → — current step (include attempt info if looping)
• · — pending
• ⚡ — auto-approved gate
• ✗ — blocked/failed

Include short result details only when useful (test counts on completed steps, attempt progress on current step, failure reason on blocked steps).

Verbose Mode Precedence

1. Executor invocation override wins — user says "run with verbose progress"
2. Workflow frontmatter — progress: verbose
3. Default — non-verbose (per-step log only, no full list)

What to Do If No Workflow Found

If no canonical docs/plans/*-workflow.md or legacy hotl-workflow*.md is found: "No workflow file found. Would you like to:

1. Create one from a design doc ($hotl:writing-plans in Codex, /hotl:write-plan in Claude Code)
2. Use a workflow template from the plugin (workflows/feature.md, workflows/bugfix.md, workflows/refactor.md)"

Source: hashgraph-online/awesome-codex-plugins → plugins/yimwoo/hotl-plugin/skills/loop-execution/SKILL.md