kaola-workflow-plan-run

name

kaola-workflow-plan-run

description

Use when executing a frozen adaptive workflow-plan.md — executes via a running-set scheduler; each frontier unit dispatched concurrently up to the fan-out cap (critical-path-first); planner-proven-disjoint (parallel_safe) write frontiers co-open in isolated legs BY DEFAULT — no operator toggles — with serial as the fallback for overlapping/uncertain writes or hosts without worktree support. Resume-safe. Mirror of commands/kaola-workflow-plan-run.md for Codex runtime.

Codex Profile Freshness Gate

On every entry or resume into this skill, before any role probe, retry, re-plan, or real dispatch, run the normal preflight gate, not --doctor. Resolve exactly one enabled installed Kaola edition from codex plugin list --json, then execute the bundled kaola-workflow-codex-preflight.js from that edition's exact marketplace/name/version cache tuple. Never search $PWD/plugins or select the lexically first cache entry:

if ! KAOLA_CODEX_PLUGIN_LIST_OUT="$(codex plugin list --json 2>&1)"; then
  printf 'profile_preflight_refused: plugin metadata unavailable: %s\n' "$KAOLA_CODEX_PLUGIN_LIST_OUT" >&2
  exit 1
fi
if ! KAOLA_CODEX_PLUGIN_META="$(node -e '
const value=JSON.parse(process.argv[1]);
const allowed=new Set(["kaola-workflow","kaola-workflow-gitlab","kaola-workflow-gitea"]);
const rows=(Array.isArray(value.installed)?value.installed:[]).filter(row => row && row.installed === true && row.enabled === true && allowed.has(row.name));
if(rows.length!==1)throw new Error(`expected exactly one enabled installed Kaola edition; got ${rows.length}`);
const row=rows[0];
for(const [label,item] of [["marketplace",row.marketplaceName],["name",row.name],["version",row.version]])if(typeof item!=="string"||item==="."||item===".."||!/^[A-Za-z0-9._-]+$/.test(item))throw new Error(`unsafe ${label}`);
if(row.pluginId!==`${row.name}@${row.marketplaceName}`)throw new Error("plugin identity mismatch");
process.stdout.write([row.marketplaceName,row.name,row.version].join("\t"));
' "$KAOLA_CODEX_PLUGIN_LIST_OUT" 2>&1)"; then
  printf 'profile_preflight_refused: invalid plugin metadata: %s\n' "$KAOLA_CODEX_PLUGIN_META" >&2
  exit 1
fi
IFS=$'\t' read -r KAOLA_CODEX_MARKETPLACE KAOLA_CODEX_PLUGIN_NAME KAOLA_CODEX_PLUGIN_VERSION <<< "$KAOLA_CODEX_PLUGIN_META"
KAOLA_CODEX_CACHE_ROOT="$HOME/.codex/plugins/cache"
if ! KAOLA_CODEX_PREFLIGHT="$(node -e '
const fs=require("fs"),path=require("path");
const [home,base,marketplace,name,version]=process.argv.slice(1);
const resolvedHome=path.resolve(home),resolvedBase=path.resolve(base);
if(resolvedBase!==path.join(resolvedHome,".codex","plugins","cache"))throw new Error("plugin cache root escapes HOME");
let cursor=resolvedHome;
const homeStat=fs.lstatSync(cursor);
if(homeStat.isSymbolicLink()||!homeStat.isDirectory())throw new Error("HOME is unsafe");
const parts=[".codex","plugins","cache",marketplace,name,version,"scripts","kaola-workflow-codex-preflight.js"];
for(let index=0;index<parts.length;index+=1){
  cursor=path.join(cursor,parts[index]);
  const stat=fs.lstatSync(cursor);
  if(stat.isSymbolicLink())throw new Error(`symlink cache component: ${cursor}`);
  if(index<parts.length-1&&!stat.isDirectory())throw new Error(`non-directory cache component: ${cursor}`);
  if(index===parts.length-1&&!stat.isFile())throw new Error(`preflight is not a regular file: ${cursor}`);
}
process.stdout.write(cursor);
' "$HOME" "$KAOLA_CODEX_CACHE_ROOT" "$KAOLA_CODEX_MARKETPLACE" "$KAOLA_CODEX_PLUGIN_NAME" "$KAOLA_CODEX_PLUGIN_VERSION" 2>&1)"; then
  printf 'profile_preflight_refused: exact active preflight unavailable: %s\n' "$KAOLA_CODEX_PREFLIGHT" >&2
  exit 1
fi
KAOLA_CODEX_PREFLIGHT_ARGS=(--project-root "$PWD" --no-autofix --json)
if [ -n "${KAOLA_CODEX_PREFLIGHT_PLAN:-}" ]; then
  KAOLA_CODEX_PREFLIGHT_ARGS+=(--plan "$KAOLA_CODEX_PREFLIGHT_PLAN")
fi
if ! KAOLA_CODEX_PREFLIGHT_OUT="$(node "$KAOLA_CODEX_PREFLIGHT" "${KAOLA_CODEX_PREFLIGHT_ARGS[@]}" 2>&1)"; then
  printf 'profile_preflight_refused: %s\n' "$KAOLA_CODEX_PREFLIGHT_OUT" >&2
  exit 1
fi
if ! KAOLA_CODEX_PREFLIGHT_STATUS="$(node -e 'const v=JSON.parse(process.argv[1]);if(typeof v.status!=="string")throw new Error("missing status");process.stdout.write(v.status)' "$KAOLA_CODEX_PREFLIGHT_OUT" 2>&1)"; then
  printf 'profile_preflight_refused: malformed preflight result: %s\n' "$KAOLA_CODEX_PREFLIGHT_STATUS" >&2
  exit 1
fi
if [ "$KAOLA_CODEX_PREFLIGHT_STATUS" != ok ]; then
  printf 'profile_preflight_refused: %s\n' "$KAOLA_CODEX_PREFLIGHT_OUT" >&2
  exit 1
fi

The exact active cache root is $HOME/.codex/plugins/cache/$KAOLA_CODEX_MARKETPLACE/$KAOLA_CODEX_PLUGIN_NAME/$KAOLA_CODEX_PLUGIN_VERSION. The base invocation is --project-root "$PWD" --no-autofix --json; the gate merges persisted config from HOME through the repository root to "$PWD". When this skill owns a frozen adaptive plan, set KAOLA_CODEX_PREFLIGHT_PLAN to that exact plan before running the block so --plan is also enforced. Continue only after exit 0 and parsed status: "ok". Exact-byte drift such as profile_bytes_mismatch is profile_preflight_refused: STOP before any agents.spawn_agent call, never record subagent-invoked, and do not relabel profile/config drift as tool unavailability or local fallback. Re-run the gate if the installed profile set changes.

Skill: kaola-workflow-plan-run (GitLab)

Adaptive executor. Runs a frozen workflow-plan.md (workflow_path: adaptive) by traversing its DAG + ## Node Ledger instead of the fixed phaseN ladder. Reads and updates kaola-workflow/{project}/workflow-state.md throughout. The plan is guarded by plan_hash; tampering is a typed refusal. Drive every node to complete or n/a, honoring the computed gates, then route to kaola-workflow-finalize.

Run subcommands with --summary for one-line output. For an opening call (open-next / open-ready / close-and-open-next), the summary line already carries the dispatch essentials: summary: ok | opened=<node-id> role=<role> task=<codex_task_name> mode=<codex_dispatch_mode> effort=<medium|xhigh|unresolved> (one opened= segment per member on a batch open; effort=unresolved is a typed codex_tier_unresolved refusal, never inheritance; the leg path is NOT in the summary line). The full envelope — every field, including dispatch.leg_path and the complete dispatch:{...} object — needs --json without --summary, or the cached .cache/<op>-envelope.json. Drill into the full envelope on result: refuse (includes operator_hint), AND — whenever running with --summary — before every dispatch: take the dispatch card from the summary line's opened= segment or from .cache/<op>-envelope.json. Never dispatch without the card in view.

In-progress re-plan control plane

This fence outranks every normal startup, mirror, scheduler, handoff, validation, and finalization route. Before any such action, read the project state and transaction status. When either reports replan_in_progress, do not mutate or replace the frozen parent workflow-plan.md. Read-only orientation must report the exact replan_phase, transaction_id, parent_plan_hash, child_plan_hash (or none), and last_cas_result; never reconstruct them from memory.

The single legal mutation while the fence is active is the edition-local re-plan resume command:

REPLAN_SCRIPT="./plugins/kaola-workflow-gitlab/scripts/kaola-gitlab-workflow-replan.js"
if [ ! -f "$REPLAN_SCRIPT" ]; then
  REPLAN_SCRIPT="$(find "$HOME/.codex/plugins/cache" -path '*/kaola-workflow-gitlab/*/scripts/kaola-gitlab-workflow-replan.js' -print -quit 2>/dev/null)"
fi
[ -n "$REPLAN_SCRIPT" ] && [ -f "$REPLAN_SCRIPT" ] || { echo "BLOCKED: kaola-gitlab-workflow-replan.js unavailable" >&2; exit 1; }
node "$REPLAN_SCRIPT" resume --project {project} --json

The installed aggregator is kaola-gitlab-workflow-replan.js. Do not run mirror/open/record/close/run-chains, ordinary adaptive handoff, claim archive, task-mirror refresh, or finalize while an intermediate phase remains. decision:ask remains advisory and never adds a pause or gate.

If resume returns replan_planner_dispatch_required, dispatch the genuine workflow-planner profile in its Re-plan dispatch mode with an isolated brief containing only the repository root, project, transaction_id, dispatch_nonce, profile identity, the exact .cache/replan-planner-packet.json path, and the packet's reason/source evidence. No role sequence, node ids, dependencies, write sets, cardinality, shape, model, or exact DAG fragment may be supplied by the orchestrator; an attempt earns planner_control_boundary_violation. The planner alone writes the seeded workflow-plan.next.md and .cache/replan-planner-attestation.json, then returns through this same resume command. Missing, stale, replayed, or mismatched dispatch proof/attestation is replan_planner_attestation_invalid; main must never synthesize either artifact.

An invalid unfrozen child uses the bounded unfrozen child-repair loop: re-dispatch the same planner with the verbatim validator errors and its own child draft, then resume. The main session never repairs the child DAG. At the retry bound, stop with the typed evidence; do not create a competing plan, restart the claim, or route to another path. A verified legacy-v1 parent follows this same transaction into a schema-2 child; legacy normal startup behavior otherwise stays unchanged.

Setup

Resolve the worktree path and $KAOLA_SCRIPTS before the first node call:

ACTIVE_WORKTREE_PATH="$(node -e "try{const fs=require('fs');const s=fs.readFileSync('kaola-workflow/' + process.env.KAOLA_PROJECT + '/workflow-state.md','utf8');const m=s.match(/^worktree_path:\\s*(.+)$/m);process.stdout.write(m?m[1].trim():'');}catch(e){}" 2>/dev/null)" || true
[ -z "$ACTIVE_WORKTREE_PATH" ] && ACTIVE_WORKTREE_PATH="$(pwd)"
KAOLA_SCRIPTS="plugins/kaola-workflow-gitlab/scripts"
if [ ! -f "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" ]; then
  KAOLA_SCRIPTS="$(dirname "$(find "$HOME/.codex/plugins/cache" -path '*/kaola-workflow-gitlab/*/scripts/kaola-gitlab-workflow-adaptive-node.js' -print -quit 2>/dev/null)")"
fi

Resolve the scripts path per command, not once — the Bash tool does not persist env vars between calls, so re-resolve $KAOLA_SCRIPTS (repeat the resolver, or hardcode the absolute path) in EVERY Bash call that needs it; a once-set value is GONE in the next call.

Then mirror the project folder into the worktree (idempotent, plan_hash-verified):

node "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" mirror-project \
  --project {project} --json

status: mirrored | exists | skipped → proceed. result: refuse → STOP and surface verbatim.

Then enter the worktree — every adaptive lifecycle call below runs from the worktree cwd:

[ -n "$ACTIVE_WORKTREE_PATH" ] && [ "$ACTIVE_WORKTREE_PATH" != "$(pwd)" ] && cd "$ACTIVE_WORKTREE_PATH"

Worktree-cwd contract: the mutating lifecycle subcommands (open-next / open-ready / record-evidence --stdin / close-and-open-next / close-node / reconcile-running-set / write-halt / clear-halt / reopen-node / revert-overflow / repair-node / route-findings) resolve the project folder — workflow-plan.md, the ## Node Ledger, .cache/{node-id}.md evidence, and the barrier baselines — cwd-relative. Run them ALL from the worktree so durable state lands where the role agents write. Running one from the main repo root while a worktree is linked refuses with worktree_authority_split (zero mutation) — cd "$ACTIVE_WORKTREE_PATH" and re-run. (orient and record-evidence --verify are read-only and exempt; mirror-project is the main→worktree copy and must run from the main root, above.)

Dispatch

Model, reasoning effort, and identity: always delegate to the base dispatch.agent_type profile (= the node's role). Every named profile omits both runtime-strength keys and therefore inherits the current parent session. The descriptor's dispatch.codex_profile_mode is inherit; its dispatch.codex_profile_tier remains declarative role metadata, and its dispatch.codex_model + dispatch.codex_reasoning_effort fields come only from a fresh parent-session proof. An unresolved tier refuses as codex_tier_unresolved; inheritance never creates a tier/profile compatibility conflict.

Before the first real node for the current parent session and installed profile-set fingerprint, require one fresh parent-equals-child inheritance proof. Spawn the probe by role identity only, bind it to the intended installed profile path and role, then inspect both parent and child session JSONL latest turn_context model and effort. Config text, spawn arguments, and parent-side descriptors are not proof. If the profile-set binding changes, either value is absent/stale, the child identity/path is wrong, or the child pair differs from the freshly re-read parent pair, refuse with codex_profile_runtime_mismatch. A parent-side encrypted output/decryption failure does not erase valid JSONL profile evidence. For real node work it is a separate transport failure handled only by the durable-result contract below: verified child-written cache evidence may continue as returned_partial; missing or invalid cache evidence cannot.

Codex collaboration transport is a hard pre-dispatch gate. In v2 task-name mode, invoke every collaboration operation through the direct agents namespace reported by preflight. Never use the server-reserved collaboration namespace and never dispatch through functions.exec or Code Mode. If preflight reports codex_v2_encrypted_transport_unsafe or codex_v2_role_transport_unsafe, refuse before spawning. Do not retry an encrypted-output decode or reserved-schema failure and do not fall back to a default role: the same transport/schema mismatch is deterministic.

For Codex v2 task-name mode (dispatch.codex_dispatch_mode: "v2-task-name"), after the proof gate passes, call the direct agents.spawn_agent tool with task_name: dispatch.codex_task_name, agent_type: dispatch.agent_type, and fork_turns: "none" on EVERY role dispatch — the dispatch card is self-contained by contract, so no role spawn ever forks the parent's history. Omit both model and reasoning_effort; the named standalone profile owns the pair. For v1 fallback ("v1-thread-id"), omit task_name, keep agent_type: dispatch.agent_type and fork_turns: "none" — the unconditional mandate applies identically to this dispatch mode — again omit model and reasoning_effort, and prefix the prompt with Node: <id> | Role: <role> | Expected model: <dispatch.codex_model> | Expected effort: <dispatch.codex_reasoning_effort> | Profile mode: <dispatch.codex_profile_mode>. Pass Working directory: ${ACTIVE_WORKTREE_PATH} to every role delegation.

Codex Join Protocol

Spawning a frontier is only half the delegation lifecycle — this protocol governs everything after spawn_agent returns: how long to wait, when (and how) to nudge a slow agent, and how to reclaim a node safely if it truly stalls. No timeout or patience value here is left to model improvisation — every number traces to the dispatch card (dispatch.wait_budget_minutes) or a named config bound. Detailed mechanics: docs/plan-run-cards/join-protocol.md.

A. Wait budget. Every dispatch card carries dispatch.wait_budget_minutes — tier-derived (e.g. 40 minutes for a reasoning-tier node, 20 for standard; an unresolved tier still resolves to a concrete role-default, never null). A running agent is NEVER interrupted before its wait budget expires. Read the budget off the card at dispatch time; never substitute an improvised patience ceiling.

B. Long-poll join loop — one wait per iteration, drain every completed member, no status probes. After dispatching a frontier, loop: call wait_agent ONCE per iteration with a LONG timeout (minutes — at or near the host's max_wait_timeout_ms), passing every outstanding agent id where the tool supports multi-id wait. On wake, call list_agents ONCE, then drain EVERY completed member before re-waiting — integrate its result, record-evidence, close-node, and, where the tool surface exposes it, close_agent (best-effort hygiene: some sessions never expose it and the harness auto-reaps completed agents, so its absence is not an error). send_message status-probe requests to a still-running agent are PROHIBITED as a liveness check — a busy agent answers at its own turn boundary, not on demand, and a probe is structurally unanswerable evidence of nothing.

C. Escalation ladder — replaces impatience-kill. Applied ONLY after the wait budget (A) has expired, each rung gated on the previous:

1. Budget expired → send a followup_task demanding the bounded deliverable now (evidence + changed-file list); this reaches a running agent at its next message boundary.
2. Grace window (~5 minutes) passes with no response → interrupt_agent, then a further followup_task asking the still-available agent for partial evidence and its changed-file list.
3. Only then: reclaim the node. Inline redo by the orchestrator is the documented LAST resort.

Record a typed delegation_outcome in the node's evidence for every delegation: completed | returned_partial | interrupted_unresponsive | interrupted_obsolete — never a free-text "it stalled so I did it myself".

Writer kill-safety. An in-place writer (shared worktree) is non-interruptible before the wait budget and the full escalation ladder above — no exception. A writer that must be interruptible belongs in an isolated parallel_safe leg instead (the existing per-leg mechanism); interrupting an isolated-leg writer discards the leg atomically, never partially. After reclaiming ANY writer (ladder step 2 or the reclaim itself), run reconcile-running-set and HONOR its verdict before re-opening the node: a writerHalt: true result means at least one departing writer's changes could not be positively confirmed inside its declared write set — do NOT re-open that node until the out-of-set paths are resolved (revert-overflow, repair-node, or a consent halt). Re-opening on a halt verdict without resolving it first is the exact halt-then-reopen laundering hole this protocol closes.

F. Frontier dispatch discipline + slot awareness. On enterBatch: true, issue every spawn_agent call for the frontier back-to-back in ONE turn, then run exactly ONE join loop (B) for the whole frontier — never one spawn-then-wait cycle per member. Width counts only RUNNING members: a finished-but-unclosed agent does not hold a concurrency seat, so do not subtract it from available width. On a spawn refusal (a thread-limit / concurrency-limit error), wait for or close a finished agent, then retry the SAME spawn ONCE — this reactive fallback is the capacity remedy, never proactive closure.

The dispatch card's frozen wait_budget_minutes value and source are authoritative. A planner_override may extend but never shorten the existing no-interrupt floor. The join loop must not interrupt or re-nudge before that floor expires; after it expires, the bounded escalation still requires a complete governed deliverable. This planner-authored floor is distinct from optimize_budget: only a metric-optimizer contract supplies the specialized optimization wall-clock source described by the metric-optimizer card.

Gate-Role Degradation Notice

The Codex Profile Freshness Gate above is authoritative for profile/config availability. Missing, stale, malformed, or shadowed project profiles are profile_preflight_refused: STOP before opening the first node; never route that drift through this degradation path. After a successful gate, determine runtime dispatch-tool availability before opening the first node and re-check if it changes mid-run. Only a genuinely unavailable agent tool or a runtime mode-refused spawn qualifies. When that dispatch capability is unavailable, post a PROMINENT run-start notice — before dispatching any node — naming every gate role the plan would otherwise have dispatched: adversarial-verifier, code-reviewer, security-reviewer.

For adversarial-verifier and code-reviewer, an inline gate reviewing its own writer-context is no gate: do NOT dispatch the gate node inline and silently record a self-issued verdict: pass. Instead route through the consent-halt valve (write-halt --reason consent) and await operator resolution before the gate node is considered satisfied. Forward roles — code-explorer, knowledge-lookup, implementer, tdd-guide, metric-optimizer, doc-updater, and security-reviewer when it runs as a forward check — may still record the documented local fallback (local-fallback-tool-unavailable) and proceed inline.

When a node runs inline under this degradation notice, announce it instead of the pre-spawn format above:

→ running {node_id} · {role} inline (…reason token…)

Loop Skeleton

1. Orient (on entry / resume)

node "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" orient \
  --project {project} --json --summary

orient is read-only: re-checks plan_hash, the ready set from next-action, any in_progress node and its .cache/{node-id}.md state, escalated_to_full / consent_halt: pending markers, and the allDone flag. Makes NO mutations.

On resume or after a context compaction, the in-progress node's re-dispatch context (goal_line + upstream_evidence) is re-derived from the cached .cache/<op>-envelope.json — the disk is authoritative; never reconstruct it from memory.

On crash/interrupt resume, read the card: docs/plan-run-cards/resume.md (covers requires_redispatch, complete-evidence crash, consent-halt clear-halt, unfrozen plan)

On plan_not_frozen / governance: docs/plan-run-cards/governance.md (covers decision:auto-run/decision:ask audit metadata; provisional run authorization; auto-run vs ask — a frozen in-grammar plan RUNS either way; typed refusal on out-of-grammar)

Announce the run once, right after this orient succeeds and before opening the first node:

plan-run orchestrator: driving {project} — {N} nodes; each role subagent will be announced at dispatch.

Substitute {project} for the project name and {N} for the total row count of ## Nodes.

2. Open next node

node "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" open-next \
  --project {project} --json --summary

Under --summary (the canonical invocation above), the printed line is summary: ok | opened=<node-id> role=<role> task=<codex_task_name> mode=<codex_dispatch_mode> effort=<medium|xhigh|unresolved> or summary: ok | allDone: true — refuse an unresolved tier; otherwise read the dispatch card straight off the opened= segment. The full envelope needs --json without --summary (or the cached .cache/open-next-envelope.json): {opened:{id,role,model,declared_write_set}, nonce, evidence_file, required_tokens, dispatch:{...}} or {allDone:true}. On allDone, run chains then route to finalize.

The fused close-and-open-next (step 4) opens every subsequent node. Re-run open-next only when no node is in_progress.

3. Dispatch the role agent

Mirror before dispatch. Apply the returned taskTransitions to the visible task list BEFORE spawning the role agent — the ledger stays authoritative; the mirror is the operator's only live view.

Every spawn parameter comes from the dispatch card. NEVER improvise a task name, omit agent_type, or drop the effort tier because the card was not in view — go get the card first (the summary line's opened= segment, or .cache/<op>-envelope.json).

Reviewer Contract Envelope, Validation, and Convergence

Resolve the frozen plan contract before dispatch. A verified already-frozen plan whose hash-covered Meta predates the version field is the explicit legacy branch: plan_schema_version: 1 and contract_version: 1. Preserve its existing token registry, verdict/findings_blocking evidence, and schema-1 journal byte-for-byte; never manufacture a schema-2 context or rewrite legacy state. Any new/mismatched/unknown version is a typed refusal.

For a schema-2 review gate, the opener's dispatch card is the sole runtime envelope. Pass every field unchanged: plan_schema_version, contract_version, behavior_contract_version, behavior_contract_hash, resolved_profile_hash, review_context_hash, review_context_path, candidate_digest, graph-derived gate_mode, logical_gate, gate_claim, gate_surface, and gate_aggregation. Never infer a second mode, assemble a context from prose, substitute a local profile, or omit a field. The role reads the canonical file at review_context_path before it examines findings and echoes the card's required identity tokens; execution_status and gate_effect remain harness-owned.

Before settling that gate, read the context's validation_obligations. For every inherited obligation, run the frozen schema-2 validation policy through the validation runner and persist its canonical receipt under .cache/validation-vectors/. Accept it only when the machine comparison finds the exact obligated command/vector identity, the current candidate_digest, and outcome pass; missing, drifted, failed, inconclusive, timed-out, signaled, or candidate-mutating execution leaves the gate open and routes validation repair. Never relabel an execution failure as an analytical indeterminate outcome.

Resolve kaola-workflow-validation-runner.js from the same installed runtime as the adaptive node; pass the exact hash-covered policy values and write one canonical JSON receipt per obligation without shell re-derivation.

Honor close-time convergence outcomes structurally. review_failed follows the authoritative journal and existing agent-owned writer selection. replan_required with review_scope_expanded or review_nonconvergent is a settled durable handoff: surface the exact packet and stop this frozen run without calling repair-node, thawing the plan, activating another epoch, or choosing topology. The harness never selects a writer or replacement DAG.

Immediately before every spawn, announce the dispatch:

→ dispatching {node_id} · {role} as subagent task "{task_name}" (model {model}, effort {effort})

{task_name} is dispatch.codex_task_name on Codex, the agent name/description on Claude, the child task label on opencode. {effort} is the dispatch-card effort on Codex and n/a on a runtime without an effort surface; it never means parent/session inheritance.

Delegate to the base role profile matching dispatch.agent_type. Apply the task-name and reasoning-effort rule above. Pass dispatch.nonce (evidence-binding token). Instruct the role to:

• Read the seeded .cache/{node-id}.md (dispatch.evidence_file) for required tokens.

• When dispatch.goal_line is present, carry it VERBATIM into the role dispatch as the node's task direction.
• When dispatch.upstream_evidence is present, instruct the role to READ each listed evidence file BEFORE starting work and record a column-0 upstream_read: <node-id> <nonce> line in its own evidence, copying the nonce from line 1 of that upstream file (never from the card — the card never carries it).
• Fill in token stubs; NEVER modify the evidence-binding: header line.
• finalize sink and main-session-gate are non-delegable — run main-session-direct. Record compliance as main-session-direct for the finalize sink node.
• Gate roles must post-dominate every code/sensitive node in the ## Node Ledger; emit verdict: pass|fail + findings_blocking: N. Run --forbidden-only for forge-touching nodes. Forge-port mirror nodes: instruct with the full accumulated root diff diff spec.
• For read-only fan-out (quorum/tally-fn/validateNodeOutput): dispatch concurrently, persist and verify evidence by the runtime contract below, close-node per member.
• FANOUT_CAP (default 4) is a runtime limit, not a planning cap. The rolling top-up re-run of open-ready (admitting a NEW member as a slot frees) drains a wider READ fan-out only; a WRITE frontier wider than FANOUT_CAP does NOT top-up into a live lane group (group membership / write_union / baseline are fixed at group formation, and write_node_exclusive fires while any member is live) — it runs as fixed group waves: the first ≤cap members form a group and run to completion (each wave paying its own synthesizer-merge + group barrier), then the next wave forms as a NEW group, so makespan is the sum of the per-wave maxima, not a rolling drain. KAOLA_FANOUT_CAP_READONLY (default 8) applies to read-only fan-out.
• Planner-proven-disjoint (parallel_safe), shared-infra, and coarse (same non-shared top-level area, exact-file-disjoint — e.g. two cross-edition antichains both under plugins/) write frontiers ALL co-open in isolated legs BY DEFAULT — no operator toggles — under the retained net (a post-dominating code-reviewer gate over the legs, no PROTECTED file in either set). Serial (max_concurrent=1) is the FALLBACK only for a genuine exact-path overlap (same file or a case-collision), a directory/glob-shaped entry that cannot prove exact-path disjointness, the retained net not holding, hosts without worktree support, or an explicit KAOLA_PARALLEL_WRITES=0 opt-out; --write-overlap-consent is parsed for frozen-plan back-compat but is VESTIGIAL at this seam — see the leg-isolation note below. opening marker + reconcile-running-set handle crash-resume. test_thrash ≥ 3: escalate via write-halt --reason test_thrash.

A main-session-gate node body never instructs authoring files — it verifies. When gate instrumentation is needed to run the check (a probe scene/test/fixture, including build wiring), an UPSTREAM writer node provisions it inside that node's own declared write set; the gate only runs what was provisioned. A gate-window fence backs this at runtime: an in-worktree out-of-band Write/Edit while a main-session-gate is open is denied by default (KAOLA_GATE_WINDOW_FENCE=0 opt-out); legal exits are provisioning via an upstream writer node, route-findings/repair-node, or write-halt --reason consent.

Write-parallelism is default-on for disjoint AND same-area (coarse) frontiers. The per-leg isolation engine is COMPLETE and live, and every exact-file-disjoint write frontier — planner-proven-disjoint (parallel_safe) siblings in different top-level areas, a shared-infra frontier in the same infra area, or a coarse frontier in the same non-shared top-level area (e.g. two cross-edition antichains both under plugins/) — co-opens as isolated parallel legs BY DEFAULT — no operator toggles, gated only on the retained net: a post-dominating code-reviewer gate over the legs, and no PROTECTED file in either set. Per-leg worktree isolation + the mandatory synthesizer reconcile are the correctness net; co-open ALWAYS provisions a dedicated leg per write sibling (group-form ⟺ legs provisioned — never the legless attribution-blind union barrier).

• Serial is the FALLBACK only for a genuine exact-path overlap (same file or a case-collision), a directory/glob-shaped entry that cannot prove exact-path disjointness, the retained net not holding (no post-dominating gate, or a PROTECTED file in either set), hosts without worktree support, or an explicit KAOLA_PARALLEL_WRITES=0 opt-out (which forces serial).
• --write-overlap-consent and write_overlap_policy are parsed for frozen-plan back-compat but are VESTIGIAL at this seam — they neither enable nor block any co-open decision here. A genuinely-overlapping (exact-path or case-collision) frontier serial-degrades regardless of consent — no cross-contamination, no silent loss. No consent flag is needed for any disjoint, shared-infra, or coarse frontier.

On open-next → gate_not_complete with a speculative gate (speculative_open_policy: auto — the freeze-time default — or consent, in plan ## Meta): docs/plan-run-cards/speculative-open.md (covers open-ready's speculative activation — automatic at auto, --speculative-consent at consent — discard-speculative, gate verdict:fail rollback)

Speculative gate-overlap is default-on (speculative_open_policy: auto — the freeze-time default; the three tiers are auto / consent / off). A node blocked only by a still-open gate opens the moment open-ready runs; --speculative-consent is a no-op at auto and required only at consent, and plain serial waiting is the DEGRADED path (the sole behavior at off). Eligibility, write-speculation safety, discard, and telemetry mechanics live in the card above.

• Write-leg dispatch discipline. Isolation is discipline-dependent, not transparent — the Agent tool has no cwd parameter and a provisioned .kw/legs/<project>/<node> leg does NOT auto-redirect a leg agent's edits. Dispatch each leg with its member's dispatch.leg_path (and dispatch.leg_branch) from the open-ready payload — every Edit/Write uses an absolute <dispatch.leg_path>/... path and every Bash uses cd "<dispatch.leg_path>" && (the load-bearing instruction in the leg brief), removing the need for a laneGroup cross-reference (the laneGroup descriptor remains for observability). The failure mode is fail-closed containment, not construction: a relative-path own-lane slip lands in the parent (invisible to the per-leg barrier, exempt from the write-lane hook) and is caught by the parent-clean fence before the merge → merge_conflict/repair, never silent cross-contamination or silent loss. Bounded thrash: after K = 2 repair nudges on an agent that keeps writing out-of-lane, escalate to a merge_conflict halt rather than looping.
• merge_conflict (write-overlap): a write-leg level whose FIRST-detection refusal — member_vacuity (a no-op leg), write_set_overflow (an overflow), or the synthesizer's octopus bail (a real same-file conflict) — survives MERGE_CONFLICT_REPAIR_LIMIT (K=3) bounded repairs. Repair each first by its own recovery (re-dispatch the leg · revert-overflow · a reasoning-class (non-lowerable floor) synthesizer agent resolves a real conflict by intent), re-running close-node; on the K-th failure escalate via write-halt --reason merge_conflict. Routed exactly like test_thrash (a schema constant the orchestrator applies — NO script counter on the adaptive path); the COMMIT-based union barrier on M, never the counter, is the fail-closed gate, so a resumed run safely re-counts from zero. RESUMABLE consent-style halt — resolve, then clear-halt --reason consent.

Evidence-persistence contract per runtime and role-kind. The authoritative inter-node handoff is always the full nonce-bound .cache/<node-id>.md artifact. A self-writing role returns only a compact summary after persisting it. On runtimes that enforce a read-only role tool manifest, the role instead returns the full body as transport for orchestrator persistence; after that write, the cache artifact is again authoritative. The mechanism is derived from each role's tool manifest — no hand-list, no per-agent guesswork:

• Any node role WITHOUT Write in its tool manifest CANNOT self-write .cache evidence — it RETURNS its deliverable for orchestrator persistence via record-evidence --stdin (below). record-evidence re-injects this node's evidence-binding: header, so persisting evidence cannot strip the header — the role MUST NOT try to add or modify it. Current roster (EXAMPLES only): read producers code-explorer, knowledge-lookup, code-architect, planner, issue-scout; plus the read gates adversarial-verifier / code-reviewer / security-reviewer.
• Any node role WITH Write in its tool manifest SELF-WRITES its .cache evidence, INCLUDING the seeded evidence-binding: header (read it from the seeded file, never alter it). Current roster (EXAMPLES only): implementer, tdd-guide, metric-optimizer, build-error-resolver, doc-updater, synthesizer.

Codex 0.144 durable-result override. Every Codex node-role profile, including a logically read-only research/review role, MUST write its FULL structured deliverable directly to the exact dispatch.evidence_file before sending its final message. For a read-only role this one seeded workflow-cache file is the sole write exception; repository/product files remain forbidden. Read the existing evidence-binding: <node-id> <nonce> header, preserve it byte-for-byte, and append/replace only the body below it. The cache body must include every required token, verdict/finding row, source, and upstream-read attestation. The final message to the main orchestrator is deliberately compact: <node-id> <role>: <verdict|outcome>; evidence=<dispatch.evidence_file> — never retransmit the full deliverable as the only durable copy.

After the child stops, validate the on-disk artifact before closing or opening a dependent node:

node "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" record-evidence \
  --project {project} --node-id {node-id} --verify --json

Only result: ok makes that artifact eligible for dispatch.upstream_evidence. If the parent-side summary is disconnected or cannot be decrypted but the child is terminal and this verification is green, read the compact outcome/verdict from the cache, record delegation_outcome: returned_partial plus transport_error: encrypted_return, and continue from the durable artifact. If verification is absent or red, leave the node open and route repair; a seed-only file never counts as completed work.

A metric-optimizer node's dispatch card carries dispatch.optimize (the frozen optimize(<node-id>) metric contract) and may override dispatch.wait_budget_minutes from the contract's budget_wallclock_minutes. Full ratchet protocol: docs/plan-run-cards/metric-optimizer.md (covers the propose/apply/regression-gate/measure/ accept-or-reject loop, the metric: <number> output contract, scoped-revert safety, stop conditions, and the five evidence tokens).

On every return, before evidence/close bookkeeping, announce the compact outcome:

← {node_id} · {role} returned: {verdict or one-line outcome}

For a returned full body that still needs parent-side persistence, record it with stdin:

node "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" record-evidence \
  --project {project} --node-id {node-id} --stdin --json

For a self-written artifact, never overwrite it with the compact summary; use record-evidence --verify as shown above.

4. Close and advance

node "$KAOLA_SCRIPTS/kaola-gitlab-workflow-adaptive-node.js" close-and-open-next \
  --project {project} --node-id {node-id} --json --summary

Enforces (in order): evidence-shape check → barrier (plan_hash re-verified, diff against baseline, post-dominate gate check, escalated_to_full: consent / typed refusal on lane overflow) → close + compliance row → selector routing → fused advance. Returns {closed:{...}, opened:{...}|null, allDone} or result: refuse.

After every close, print the required progress line:

{node-id} → complete; opened: {next-id|—}

On result: ok + opened: dispatch the next node (step 3). On allDone: true: run chains then route to Finalization. On opened: null + allDone: false (the typed reason: 'frontier_blocked' signal): do NOT park silently. Deterministically re-run orient then open-next / open-ready to re-open the recomputed frontier, draining toward allDone WITHOUT operator prompting. Cap the re-orient at a small bound (e.g. 3 consecutive frontier_blocked cycles with no progress) before escalating with stop+ask — a blocked-but-not-done frontier must never silently stall the run.

Surface the narrowed barrier reason VERBATIM. On a barrier_failed / close-node refusal the ACTIONABLE narrowed reason (write_set_overflow / write_set_granularity / lockfile_write / …) and the offending paths are now on the top-level reason / outOfAllow fields. Surface them VERBATIM — print the full reason / operator_hint / outOfAllow — never route the refusal through a lossy JSON-summary helper that truncates it.

Authoritative failed-review repair protocol. review_failed is a settled failed transaction. Read the authoritative review-attempts.json attempt (attempt_id, logical_gate, outcome, reason, route_candidates, lifecycle_settled, repair, and consumed_by), choose the writer as an agent decision from the frozen DAG and canonical ownership_candidates, then invoke repair-node --attempt-id {attempt_id} --node-id {agent-selected-writer}. The harness never selects a repair owner and never rewrites the DAG. On retry or reconciliation, reread the journal and resume the same attempt; treat findings-route.json only as a regenerable projection, never as authority. repair_requires_replan is a zero-mutation refusal; an unresolved attempt makes reopen-node refuse with review_attempt_unresolved. Five consumed repairs are allowed per canonical logical gate; the sixth returns repair_limit_reached. Zero candidates and multiple candidates leave owning_node: null; multiple owners never imply selection.

On barrier refusal / route-findings result: docs/plan-run-cards/repair-routing.md (covers write_set_granularity / write_set_overflow / sensitive_write_unreviewed / foreign_archive / unattributed_write; revert-overflow vs repair-node; halt for consent; write-halt --reason consent; escalated_to_full: consent / escalated_to_full: security; triage / proposed_repair object; plan-repair via --freeze; --forbidden-only check)

On reopening a complete node: docs/plan-run-cards/reopen-complete-node.md (covers reopen-node for Finalization-surfaced failures; would_orphan_in_progress refusal; mid-gate repair; read-only fan-out quorum bounded controller; LOOP_CAP / dry_streak; test_thrash escalation; validateNodeOutput absent-but-never-script-enforced note)

Dispatch fidelity — concurrent dispatch is the DEFAULT, not an option. open-next (and orient / close-and-open-next) return enterBatch: true + a frontier: [...] whenever the planner authored an INDEPENDENT frontier of width ≥2 — that is authored parallelism, and you MUST run it as authored. On enterBatch: true: run open-ready (it marks the whole frontier in_progress), then dispatch the returned nodes' role agents in ONE assistant message — multiple Agent calls in a single turn. The single-message dispatch is the only thing that yields real concurrency; dispatching one agent per turn is itself a serial barrier and silently serializes a frontier the planner authored as parallel. Do NOT open-next-then-single-dispatch a ≥2 frontier (the script now refuses to single-open it). Width stays the planner's scope-driven call: a width-1 frontier or a dependency chain returns NO enterBatch and runs serially (the normal single-dispatch path) — never force a minimum width, a "default to ≥2," or a "prefer wide" posture.

On enterBatch: true the concurrent one-message dispatch above is the DEFAULT (not optional); this card covers only the running-set MECHANICS (open-ready / close-node / reconcile-running-set; write-role lane-group co-open in isolated legs; crash repair) for each frontier unit:

frontier unit

docs/plan-run-cards/frontier-batch.md (covers open-ready / close-node / reconcile-running-set; write-role lane-group co-open in isolated legs, the synthesizer + group barrier close; the opening crash-safe marker; FANOUT_CAP vs KAOLA_FANOUT_CAP_READONLY)

5. All done

When allDone: true, detect the repo type and run the terminal validation appropriate to that repo, then delegate to kaola-workflow-finalize {project}. Finalization's sink step owns its own crash-resume journals (sink-receipt.json / sink-fallback.json) and disposes of them itself at terminal success; if a stray one turns up on a later clean and synced check, delete it — never commit it.

Self-host (npm) — package.json declares test:kaola-workflow:* scripts: Run run-chains.js with --project {project}.

Invoke run-chains with --project {project}. Always pass --project {project} to the run-chains script so its receipt lands at kaola-workflow/{project}/.cache/chain-receipt.json where Finalization's --finalize-check reads it — do NOT rely on cwd to locate the receipt. The self-host chain receipt stamp is the last action before Finalization: run it only after every code change and every test-consumed prose/doc update for the final candidate has landed. If any code-relevant or chain-asserted doc changes after the stamp, the receipt is stale and chains_stale requires a full re-run of the gated runner; do not patch or hand-edit the receipt. Workflow state and inert, non-test-consumed docs are validation-invisible and do not stale the receipt.

Consumer (non-npm) repo — no test:kaola-workflow:* scripts: Do NOT invoke run-chains.js (it can only return chains_config_missing in a consumer repo). Instead, run the plan's ## Meta validation_command and record the result in kaola-workflow/{project}/.cache/final-validation.md with a column-0 verdict: pass. Finalization's --finalize-check auto-detects consumer mode (absence of the test:kaola-workflow:* scripts) and gates on final-validation.md. If an unchanged terminal change-gate validation run already covers the final candidate, you may cite that evidence instead of rerunning. Record column-0 lines for verdict: pass, source: cited:<node-id>, validated_command, validated_at_head, and reuse_boundary; if there is any doubt about the boundary, run the command. After recording the verdict (and the citation fields when citing), bind the evidence to the exact candidate it validated: run the plan-validator ($KAOLA_SCRIPTS/…-plan-validator.js) over kaola-workflow/{project}/workflow-plan.md with --candidate-hash --json and record the emitted value as a column-0 validated_candidate_hash: line in final-validation.md. Compute it LAST — after every file the validation covered has landed; any later relevant edit stales the binding and --finalize-check refuses final_validation_stale (re-run the validation command and re-record with a fresh hash — never hand-patch the hash), while a missing line refuses final_validation_unbound. Workflow state and inert, non-test-consumed docs are validation-invisible and do not stale the binding; the gate compares the recorded hash to a recomputation and never re-runs tests.

Run-Gap Manual Seeding. When the orchestrator observes a run gap the automated scanners cannot see — transient tool noise, a manual retry, an environmental flake — append a gap: <class> — <text> line to kaola-workflow/{project}/.cache/run-gaps-manual.md immediately, BEFORE Finalization's gap sweep runs, so the mapping to noise:/filed: is machine-checked, never vacuous. A ## Run gaps entry with no matching seeded/scanned source refuses observed_gap_unseeded at Finalization.

Validation De-Duplication

Avoid redundant runs: each node validates only its affected scope; cite a prior node's passing evidence (.cache/{node-id}.md) when no relevant file changed after it, stating the exact reuse boundary (never a false No files changed absolute) — a consumer citation still needs a fresh validated_candidate_hash:. Run the full chains ONCE here at All-done using the ## Meta validation_command (the record-once discipline), not per node.