@mmerterden/multi-agent-pipeline 18.0.0 → 19.0.0

package/pipeline/scripts/gen-ref-toc.mjs

@@ -13,7 +13,7 @@
 //
 // Section level is DERIVED, not assumed. These files are not consistent: some
 // open at `##` and section at `###` (keychain.md), others open at `###` and
-// section at `##` (phases/phase-3-dev.md). The rule is the shallowest heading
+// section at `##` (phases/phase-2-dev.md). The rule is the shallowest heading
 // depth that occurs more than once - a depth used exactly once is a title, not
 // a section level.
 //

package/pipeline/scripts/graph-report.mjs

@@ -3,7 +3,7 @@
 /**
  * @file graph-report.mjs  -  GRAPH_REPORT.md, the human-and-model readable summary.
  *
- * Phase 1 reads this before deciding how wide to explore, and Phase 7 ships it
+ * Phase 1 reads this before deciding how wide to explore, and Phase 5 ships it
  * next to the knowledge base. Sections are ordered by how much they narrow a
  * search: hubs first (the types everything touches), then the module map, then
  * the long tail that is usually safe to ignore.

package/pipeline/scripts/jira-attach.sh

@@ -5,7 +5,7 @@
 #
 # Two callers need this and they used to carry their own curl: the issue
 # generator (`generate-issue.md` step 12) and visual evidence (Phase 3 / Phase 5
-# captures rendered into the Phase 7 Jira comment). Two copies of an upload is
+# captures rendered into the Phase 5 Jira comment). Two copies of an upload is
 # two places for the `X-Atlassian-Token: no-check` header to go missing, which
 # fails as a 403 that reads like an auth problem.
 #

package/pipeline/scripts/learn-from-transcripts.mjs

@@ -4,7 +4,7 @@
  *
  * WHY THIS EXISTS
  *
- * Phase 7 asks a subagent to read the run log and propose a few memory entries.
+ * Phase 5 asks a subagent to read the run log and propose a few memory entries.
  * That works for what the model noticed, and it is structurally blind to what it
  * did not: the same wrong path tried in three sessions, the command that fails
  * every time until its `--flag` variant is used, the directory that never has

package/pipeline/scripts/learning-curve.mjs

@@ -87,9 +87,9 @@ for (const e of events) {
   }
   if (e.ts && (!t.firstTs || e.ts < t.firstTs)) t.firstTs = e.ts;
   const d = e.details || {};
-  // phase-4-review.md's actual review.completed emission (raw_count=/
+  // phase-3-review.md's actual review.completed emission (raw_count=/
   // accepted=/deferred=/rejected=/approved=/duration_ms=) carries no
-  // review_cycles field - that name only appears on Phase 7's task.completed
+  // review_cycles field - that name only appears on Phase 5's task.completed
   // summary event, which this never matched. Phase 4 re-runs (and logs its
   // own review.completed) once per rework cycle, so the cycle count for a
   // task is how many review.completed events share its task_id - the same

package/pipeline/scripts/log-metric.sh

@@ -6,11 +6,19 @@
 #   {
 #     "ts":      "<ISO 8601 timestamp>",
 #     "task_id": "<jira id, github issue, or run id>",
-#     "phase":   "<0..7>",
+#     "phase":   "<0..5>",
 #     "event":   "<event name>",
-#     "details": { ...arbitrary key/value pairs... }
+#     "details": { ...arbitrary key/value pairs... },
+#     "phaseSchema": 2
 #   }
 #
+# phaseSchema names the phase vocabulary the line was written under, because
+# this file is append-only and v19.0.0 renumbered the phases. A line with
+# phase 3 means Dev under schema 1 and Review under schema 2. Lines with no
+# phaseSchema field predate v19.0.0 and are schema 1. Aggregators pick their
+# name table from this field; without it the two vocabularies are
+# indistinguishable and every historical roll-up is quietly wrong.
+#
 # Usage:
 #   log-metric.sh <task_id> <phase> <event> [key=value ...]
 #
@@ -18,7 +26,7 @@
 #   log-metric.sh PROJ-1234 4 review.completed raw_count=8 accepted=3 deferred=2 rejected=3 approved=true duration_ms=42000
 #   log-metric.sh PROJ-1234 4 triage.edge_case case=high-rejection-rate
 #   log-metric.sh PROJ-1234 3 rework.started iteration=2 accepted_blocking=2
-#   log-metric.sh #316 7 task.completed phases=8 review_cycles=2 lang=tr
+#   log-metric.sh #316 5 task.completed phases=6 review_cycles=2 lang=tr
 #
 # Atomic append: each call uses `>> "$file"` which the kernel guarantees
 # atomic for writes ≤ PIPE_BUF (4 KB on macOS/Linux). One JSONL line is well
@@ -64,8 +72,13 @@ for kv in "$@"; do
 done
 DETAILS="${DETAILS}}"
 
+# Read from the phase contract rather than hardcoding: the next renumbering
+# bumps one JSON field instead of hunting for a literal in a shell script.
+PHASE_SCHEMA="$(node -e 'const c=require("'"$(cd "$(dirname "$0")" && pwd)"'/../schemas/phases.json"); process.stdout.write(String(c.phaseSchema))' 2>/dev/null)"
+[ -n "$PHASE_SCHEMA" ] || PHASE_SCHEMA=2
+
 TS="$(date -u +"%Y-%m-%dT%H:%M:%SZ")"
-LINE="{\"ts\":\"${TS}\",\"task_id\":\"${TASK_ID}\",\"phase\":\"${PHASE}\",\"event\":\"${EVENT}\",\"details\":${DETAILS}}"
+LINE="{\"ts\":\"${TS}\",\"task_id\":\"${TASK_ID}\",\"phase\":\"${PHASE}\",\"event\":\"${EVENT}\",\"details\":${DETAILS},\"phaseSchema\":${PHASE_SCHEMA}}"
 
 # Atomic append (one line, well under PIPE_BUF).
 echo "$LINE" >> "$METRICS_FILE" 2>/dev/null || {

package/pipeline/scripts/memory-save.sh

@@ -2,7 +2,7 @@
 # memory-save.sh  -  v6.2.E
 #
 # Writes a single per-repo memory file and updates the MEMORY.md index.
-# Called by Phase 7's synthesis step (or any user skill that wants to
+# Called by Phase 5's synthesis step (or any user skill that wants to
 # persist a finding). Opt-in via prefs.global.perRepoMemory.
 #
 # Usage:

package/pipeline/scripts/migrate-prefs.mjs

@@ -18,9 +18,26 @@ import { runMain } from "../lib/fatal.mjs";
 
 const DEFAULT_FILE = path.join(os.homedir(), ".claude", "multi-agent-preferences.json");
 
-// Single source of truth for the migration target version. Referenced by the
-// migrate() bump chain AND every user-facing message, so they cannot drift.
-const TARGET_VERSION = "2.6.0";
+// The migration target, read from the schema rather than written here. It was
+// a literal, and a literal is how it came to say 2.6.0 while the schema had
+// moved to 2.7.0 - the runner then reported every current file as needing a
+// migration it had already had. Referenced by the migrate() bump chain AND
+// every user-facing message, so they cannot drift from each other either.
+const TARGET_VERSION = (() => {
+  const schemaPath = path.join(
+    path.dirname(fileURLToPath(import.meta.url)),
+    "..",
+    "schemas",
+    "prefs.schema.json",
+  );
+  const sv = JSON.parse(fs.readFileSync(schemaPath, "utf-8"))?.properties?.schemaVersion;
+  const v = sv?.const ?? sv?.enum?.at(-1);
+  if (!v) {
+    console.error("migrate-prefs: prefs.schema.json declares no schemaVersion");
+    process.exit(1);
+  }
+  return v;
+})();
 
 const USAGE = "Usage: migrate-prefs.mjs [--dry-run] [--file <path>]";
 
@@ -327,7 +344,7 @@ function migrate(prefs) {
     out.global.routines = [];
     changes.push("added empty routines (v2.4.0 user-defined /multi-agent:save routines)");
   }
-  // v2.5.0 / v14.0.0: Phase 4 Review entered the --dev phase sets and Step 1.78
+  // v2.5.0 / v14.0.0: Phase 3 Review entered the --dev phase sets and Step 1.78
   // (criteria resolution) landed. Default false because only Swift plus the two
   // store-compliance catalogs ship a scoped registry today, so blocking on a
   // coverage gap would stop every Kotlin/Python/Node/ObjC run from day one. The
@@ -488,7 +505,7 @@ function migrateReportChannels(out, changes) {
   if (typeof g.autopilotReportTimeoutSeconds !== "number") {
     g.autopilotReportTimeoutSeconds = 1800;
     changes.push(
-      "added autopilotReportTimeoutSeconds: 1800 (30 min)  -  Phase 7 always-pause timeout",
+      "added autopilotReportTimeoutSeconds: 1800 (30 min)  -  Phase 5 always-pause timeout",
     );
   }
 

package/pipeline/scripts/phase-banner.sh

@@ -12,9 +12,9 @@
 #   phase-banner.sh label <phase>                               (print canonical name)
 #
 # Examples:
-#   phase-banner.sh start 4 "Review" "deterministic gates + parallel + triage"
-#   phase-banner.sh start 4 auto                   # resolves to "Review" (en) or "İnceleme" (tr)
-#   PHASE_LANG=tr phase-banner.sh start 4 auto     # "Faz 4: İnceleme"
+#   phase-banner.sh start 3 "Review" "parallel + triage + user test"
+#   phase-banner.sh start 3 auto                   # resolves to "Review" (en) or "İnceleme" (tr)
+#   PHASE_LANG=tr phase-banner.sh start 3 auto     # "Faz 3: İnceleme"
 #
 # Status values for `end`: done | failed | skipped
 #
@@ -32,28 +32,28 @@ set -uo pipefail
 
 PHASE_LANG="${PHASE_LANG:-en}"
 
-# Canonical phase label table. Keep synced with
-# pipeline/multi-agent-refs/phases.md and the table in
-# pipeline/skills/shared/multi-agent/SKILL.md.
+# Canonical phase label table. The English half is checked against
+# pipeline/schemas/phases.json by smoke-phase-contract.sh, so it cannot drift
+# from the contract in silence the way it did through the eight-phase era. It
+# stays a case statement rather than a jq read because this runs on every
+# banner and the pipeline's zero-dependency rule (ADR-0004) applies to the hot
+# path. The Turkish half has no contract entry - labels are translations, not
+# identifiers - and is kept beside it.
 phase_label() {
   local phase="$1" lang="${2:-en}"
   case "$lang:$phase" in
     en:0) echo "Init" ;;
-    en:1) echo "Analysis" ;;
-    en:2) echo "Planning" ;;
-    en:3) echo "Dev" ;;
-    en:4) echo "Review" ;;
-    en:5) echo "Test" ;;
-    en:6) echo "Commit & PR" ;;
-    en:7) echo "Report" ;;
+    en:1) echo "Plan" ;;
+    en:2) echo "Dev" ;;
+    en:3) echo "Review" ;;
+    en:4) echo "Commit & PR" ;;
+    en:5) echo "Report" ;;
     tr:0) echo "Başlangıç" ;;
-    tr:1) echo "Analiz" ;;
-    tr:2) echo "Planlama" ;;
-    tr:3) echo "Geliştirme" ;;
-    tr:4) echo "İnceleme" ;;
-    tr:5) echo "Kullanıcı Testi" ;;
-    tr:6) echo "Commit & PR" ;;
-    tr:7) echo "Rapor" ;;
+    tr:1) echo "Plan" ;;
+    tr:2) echo "Geliştirme" ;;
+    tr:3) echo "İnceleme" ;;
+    tr:4) echo "Commit & PR" ;;
+    tr:5) echo "Rapor" ;;
     *) echo "" ;;
   esac
 }

package/pipeline/scripts/phase-tracker.sh

@@ -43,7 +43,7 @@
 #
 # Examples:
 #   phase-tracker.sh init "TASK-123"
-#   for p in 0:Init 1:Analysis 2:Planning 3:Dev 4:Review 5:Test 6:Commit 7:Report; do
+#   for p in 0:Init 1:Plan 2:Dev 3:Review 4:Commit 5:Report; do
 #     phase-tracker.sh add "${p%%:*}" "${p#*:}"
 #   done
 #   phase-tracker.sh update 0 in_progress
@@ -141,7 +141,7 @@ need_jq() {
 # Live usage ping (best-effort). Emits one per-phase update to the private
 # dashboard via usage-report.mjs, always status=running so per-phase updates
 # never fold the run into rollup counters - the terminal fold comes only from
-# the Phase 7 / halt emit that reads the real run status. The emitter no-ops
+# the Phase 5 / halt emit that reads the real run status. The emitter no-ops
 # unless prefs.global.usageLog.enabled; detached so it never blocks a boundary.
 usage_live_ping() {
   # Smoke runs exercise this script's state handling, never the live dashboard -
@@ -461,7 +461,7 @@ save_state() {
 
 # --- read-modify-write lock ---------------------------------------------------
 # tmp+rename makes each save atomic, but two concurrent invocations (e.g. two
-# parallel Phase 4 reviewers reporting `tokens`) both load the same state and
+# parallel Phase 3 reviewers reporting `tokens`) both load the same state and
 # the second save silently drops the first delta. Guard every load->save
 # sequence with a portable mkdir spinlock (macOS bash 3.2: no flock builtin).
 # The wait is bounded and the lock FAILS OPEN with a warning so a stuck lock
@@ -612,7 +612,7 @@ tracker_next_hint() {
     # Not every session carries the task tools: Claude Code provides them by
     # default only up to Opus 4.7 / Sonnet 4.6, a default that landed in
     # v2.1.268. Naming the fallback on the same line is what keeps a newer model
-    # from advancing eight phases in silence.
+    # from advancing six phases in silence.
     # `subjects` now carries Phase 2's plan steps as indented rows, and
     # update_plan takes the whole list anyway, so re-reading it is what puts
     # those steps on the Codex plan without a second mechanism. Codex has no
@@ -643,7 +643,7 @@ format_span() {
 }
 
 # The end-of-run report: what the pipeline spent, phase by phase, and what it
-# has to say about phases it could not price. Printed by Phase 7 next to the
+# has to say about phases it could not price. Printed by Phase 5 next to the
 # work summary, which covers what actually changed on disk.
 report() {
   need_jq
@@ -800,7 +800,7 @@ subjects() {
 #
 # Creation order is the ONLY ordering the native widget has - it renders by the
 # order tiles were made, not by any id inside them - so a plan that arrives at
-# Phase 2 cannot simply be appended: its rows would land after Phase 7. The
+# Phase 2 cannot simply be appended: its rows would land after Phase 5. The
 # answer is a rebuild at that one boundary, which is the same thing `:resume`
 # already does for a different reason.
 tiles_script() {
@@ -811,7 +811,7 @@ tiles_script() {
   has_subs=$(echo "$state" | jq '[.phases[]?.subs[]?] | length')
 
   # A list carrying sub-steps has to be rebuilt whole, so the rebuild wins over a
-  # narrowed batch: appending to it would land the new rows after Phase 7.
+  # narrowed batch: appending to it would land the new rows after Phase 5.
   [ "${has_subs:-0}" -gt 0 ] && new_only=""
 
   if [ -n "$new_only" ]; then

package/pipeline/scripts/plan-coverage-gate.mjs

@@ -4,7 +4,7 @@
 // Phase 4 answers "is what changed correct" and Step 1.45 answers "did every
 // planned TEST land". Nothing answered "did every planned TASK land". The
 // criteria manifest's denominator is rule IDs, not plan steps, and the only
-// place a step's status surfaced was render-work-summary.sh at Phase 7 - a
+// place a step's status surfaced was render-work-summary.sh at Phase 5 - a
 // report, printed after the commit. So a plan with seven steps could ship five
 // and read as done.
 //
@@ -116,7 +116,7 @@ if (isMain) {
         "",
         "Fails when a plan step never reached a terminal status, when a skip or",
         "failure carries no reason, or when an analysis Section 14 `Add new` file",
-        "is missing from the tree. Run before Phase 6 commit.",
+        "is missing from the tree. Run before Phase 4 commit.",
         "",
       ].join("\n"),
     );

package/pipeline/scripts/render-agent-log-cost.sh

@@ -7,7 +7,7 @@
 #
 # This is the agent-log counterpart to render-cost-summary.sh (which
 # targets PR/Jira channel bodies). The agent-log version is rendered
-# unconditionally on every Phase 7 run; the channels version is opt-in
+# unconditionally on every Phase 5 run; the channels version is opt-in
 # via prefs.global.reportContent.costSummary.
 #
 # Usage:

package/pipeline/scripts/render-work-summary.sh

@@ -2,7 +2,7 @@
 # render-work-summary.sh  -  v7.1.0
 #
 # Reads agent-state.json + phase-tracker.json + git diff, emits a concise
-# "### Work Summary" markdown block for Phase 7 channels dispatch. Intended
+# "### Work Summary" markdown block for Phase 5 channels dispatch. Intended
 # as an executive-summary companion to the existing Normal Analysis /
 # Technical Details / Test Scenarios sections  -  gives Jira/Confluence
 # readers and PR reviewers a single-screen answer to "what actually
@@ -85,7 +85,7 @@ if [ -n "$WORKTREE" ]; then
   [ -f "$WORKTREE/triage-output.json" ]  && TRIAGE_FILE="$WORKTREE/triage-output.json"
 fi
 
-# Fall back to the salvaged artefacts when the worktree is gone. Phase 6 removes it
+# Fall back to the salvaged artefacts when the worktree is gone. Phase 4 removes it
 # once the PR is open (worktree-finalize), and this script used to resolve state
 # ONLY from the worktree - so it exited 2 and the whole Work Summary silently
 # disappeared from the PR body and the Jira comment. render-agent-log-cost.sh has
@@ -203,7 +203,7 @@ total_add=0; total_del=0
 # repository check. Testing -d "$WORKTREE/.git" would be wrong regardless: in a
 # linked worktree .git is a file, not a directory.
 # Prefer the worktree; fall back to the project root with the branch by name. A
-# finalized task (Phase 6 removed the worktree after the PR) has no worktree but
+# finalized task (Phase 4 removed the worktree after the PR) has no worktree but
 # the branch is still local, so without this the Changed-files section vanished
 # from the PR body and the Jira comment even though the data was right there.
 DIFF_IN=""; DIFF_TIP=""

package/pipeline/scripts/review-file-filter.mjs

@@ -4,7 +4,7 @@
  * @file review-file-filter.mjs  -  decide what the reviewers are asked to read.
  *
  * Phase 4 has a size cap and no exclusion list. When the diff exceeds the
- * budget the cap truncates the LARGEST files first (`phase-4-review.md`, Step
+ * budget the cap truncates the LARGEST files first (`phase-3-review.md`, Step
  * 1.9), so a regenerated lockfile or a snapshot dump does not merely waste
  * tokens - it is the thing that survives while real code is cut. The cheapest
  * fix is to decide what is worth reading before the cap decides what fits.

package/pipeline/scripts/run-aggregator.mjs

@@ -39,6 +39,7 @@ import { join, dirname } from "node:path";
 import { fileURLToPath } from "node:url";
 import { costUsd } from "./_cost.mjs";
 import { resolveRunFile, canonicalRunDir } from "./_run-paths.mjs";
+import { toCurrentPhase, CURRENT_SCHEMA } from "../lib/phase-schema.mjs";
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 
@@ -157,16 +158,22 @@ function computeCost(modelKey, tokensIn, tokensOut, costTable) {
 }
 
 /**
- * Heuristic: phase 1/2 use Opus (per personas + dev-mode docs); phase 3 uses
- * Sonnet by default unless the run is a Short pipeline (Opus). Phase 4 review
- * is multi-model  -  span events carry per-reviewer model when OTel is on.
+ * Heuristic: Plan (1) uses Opus per personas + dev-mode docs; Dev (2) uses
+ * Sonnet by default unless the run is a Short pipeline (Opus); Review (3)
+ * triage is Opus and the reviewer breakdown comes from spans.
  *
  * This is a fallback when phase-tracker doesn't record the model used. Live
  * runs that emit OTel spans get per-call model attribution from the spans.
+ *
+ * The phase number is normalised first. The old body compared against the
+ * literals "1", "2" and "4", which under the six-phase contract are Plan, Dev
+ * and Commit - so a pre-v19 row and a post-v19 row would have been priced
+ * against different phases with no error anywhere.
  */
-function inferModelForPhase(phaseId) {
-  if (phaseId === "1" || phaseId === "2") return "opus";
-  if (phaseId === "4") return "opus"; // triage is opus; reviewer breakdown comes from spans
+function inferModelForPhase(phaseId, schema = CURRENT_SCHEMA) {
+  const p = toCurrentPhase(phaseId, schema);
+  if (p === 1) return "opus";
+  if (p === 3) return "opus";
   return "sonnet";
 }
 

package/pipeline/scripts/run-metrics.mjs

@@ -24,7 +24,7 @@
  *   consensusVerdict                      - unanimous-* / split / unverified
  *   buildPassed                           - per-repo build outcome
  *   diff.filesTouched / locAdded /        - size of the change, from state.diffRisk
- *   locRemoved / maxScore                   (Phase 4 Step 1.75); null when the
+ *   locRemoved / maxScore                   (Phase 3 Step 1.75); null when the
  *                                           run predates the persisted totals
  *   reviewDelta.stillPresentFinal /       - cross-round classification of the
  *   resolvedTotal / tripped                 last iteration (state.reviewIterations[].delta)

package/pipeline/scripts/runs-index.mjs

@@ -42,6 +42,16 @@ import { listRuns, logsRoot, resolveRunDir, taskIdVariants } from "./_run-paths.
 import { runMain } from "../lib/fatal.mjs";
 import { invokedDirectly } from "../lib/invoked-directly.mjs";
 
+/**
+ * The phase from which a run counts as "waiting on you", read from the phase
+ * contract rather than written here. This threshold moved once already (it was
+ * 6 under the eight-phase contract, it is 4 under six) and nothing connected it
+ * to the renumbering, so it would have silently regrouped every run.
+ */
+const PHASE_WAITING_FROM = JSON.parse(
+  readFileSync(new URL("../schemas/phases.json", import.meta.url), "utf8"),
+).thresholds.waitingFromPhase;
+
 const GROUPS = {
   waiting: "Waiting on you",
   stopped: "Stopped mid-development",
@@ -166,7 +176,7 @@ function groupOf(state) {
   if (status === "awaiting_input" || status === "awaiting-user-test-main-checkout")
     return "waiting";
   if (prUrl) return "waiting";
-  if (Number.isFinite(phase) && phase >= 6) return "waiting";
+  if (Number.isFinite(phase) && phase >= PHASE_WAITING_FROM) return "waiting";
   if (Number.isFinite(phase) && phase === 0) return "question";
   return "stopped";
 }

package/pipeline/scripts/smoke-cross-cli-behavior.sh

@@ -204,21 +204,21 @@ echo "→ reviewer-count contract (Claude=3, Copilot=3, Codex=3)"
 # a broken installation. _smoke-root.sh handles both layouts.
 # shellcheck source=pipeline/scripts/_smoke-root.sh
 . "$(dirname "${BASH_SOURCE[0]}")/_smoke-root.sh"
-P4="${MA_REFS:+$MA_REFS/phases/phase-4-review.md}"
+P4="${MA_REFS:+$MA_REFS/phases/phase-3-review.md}"
 REVSCHEMA="${MA_SCHEMAS:+$MA_SCHEMAS/reviewer-output.schema.json}"
 TRSCHEMA="${MA_SCHEMAS:+$MA_SCHEMAS/triage-output.schema.json}"
 
 if [ -z "$P4" ] || [ ! -f "$P4" ]; then
-  echo "  ↷ SKIP: phase-4-review.md not present in this $MA_LAYOUT layout"
+  echo "  ↷ SKIP: phase-3-review.md not present in this $MA_LAYOUT layout"
 # Two independent statements, because they can drift apart: the count sentence is
 # the contract, and the matrix is what a reader dispatches from. Three regexes over
 # overlapping prose used to stand in for this and let the count sentence 300 lines
 # further down go stale for a whole release without failing.
 elif grep -qF "Claude Code 3, Copilot CLI 3, Codex CLI 3" "$P4" \
   && grep -qE '^\| Reviewer 3 .*\|.*\|.*\|.*\|' "$P4"; then
-  pass "phase-4-review declares Claude=3 / Copilot=3 / Codex=3 reviewers, and the matrix has all three host columns"
+  pass "phase-3-review declares Claude=3 / Copilot=3 / Codex=3 reviewers, and the matrix has all three host columns"
 else
-  fail "phase-4-review does not declare the CLI-aware reviewer count for all three hosts"
+  fail "phase-3-review does not declare the CLI-aware reviewer count for all three hosts"
 fi
 
 # The two Codex constraints are silent-failure shaped, so the contract has to name
@@ -226,9 +226,9 @@ fi
 # 4-slot ceiling (orchestrator included) is why the count is 3 and not more.
 if [ -n "$P4" ] && [ -f "$P4" ]; then
   if grep -q 'fork_turns' "$P4" && grep -qiE "concurrency|slots" "$P4"; then
-    pass "phase-4-review documents the fork_turns override rule + the concurrency ceiling"
+    pass "phase-3-review documents the fork_turns override rule + the concurrency ceiling"
   else
-    fail "phase-4-review must document fork_turns and the Codex concurrency ceiling"
+    fail "phase-3-review must document fork_turns and the Codex concurrency ceiling"
   fi
 fi
 

package/pipeline/scripts/smoke-schema-validation.sh

@@ -24,6 +24,26 @@ TEMPLATE="$SMOKE_DIR/../preferences-template.json"
 [ -f "$TEMPLATE" ] || TEMPLATE="$HOME/multi-agent-pipeline/pipeline/preferences-template.json"
 LIVE_PREFS="$HOME/.claude/multi-agent-preferences.json"
 
+# The migration target, derived the way migrate-prefs.mjs derives it: the last
+# entry of the schema's schemaVersion enum, read from the schema that sits
+# beside the migrator being checked. It used to be grepped as a
+# `TARGET_VERSION = "x.y.z"` literal out of the migrator, and when that literal
+# was replaced by the schema read - precisely because a literal had drifted a
+# minor behind - the grep started matching nothing and both checks that depend
+# on it reported "no reference point". A gate that reads the source of truth
+# cannot go stale against it; a gate that reads a transcription of it can.
+migration_target() {
+  local schema="$1/../schemas/prefs.schema.json"
+  [ -f "$schema" ] || schema="$PREFS_SCHEMA"
+  [ -f "$schema" ] || return 1
+  node -e "
+    const sv = JSON.parse(require('fs').readFileSync('$schema','utf8'))?.properties?.schemaVersion;
+    const v = sv?.const ?? sv?.enum?.at(-1);
+    if (!v) process.exit(1);
+    process.stdout.write(v);
+  " 2>/dev/null
+}
+
 # ──────────────────────────────────────────────────────────────────────────
 echo "→ 1. Schema files parse as JSON"
 for f in "$PREFS_SCHEMA" "$STATE_SCHEMA"; do
@@ -145,10 +165,9 @@ if [ -f "$TEMPLATE" ]; then
   # every old entry in the migrator's accepted set became load-bearing purely to
   # rescue the template this gate was holding back. A template behind the target
   # is a defect, not the expected shape.
-  TEMPLATE_TARGET=$(grep -oE 'TARGET_VERSION = "[0-9.]+"' "$SMOKE_DIR/migrate-prefs.mjs" 2>/dev/null \
-    | grep -oE '[0-9]+\.[0-9]+\.[0-9]+')
+  TEMPLATE_TARGET=$(migration_target "$SMOKE_DIR")
   if [ -z "$TEMPLATE_TARGET" ]; then
-    fail "cannot read TARGET_VERSION from migrate-prefs.mjs  -  the check has no reference point"
+    fail "cannot derive the migration target from prefs.schema.json  -  the check has no reference point"
   elif [ "$TVER" = "$TEMPLATE_TARGET" ]; then
     pass "template schemaVersion: $TVER (at the migration target)"
   else
@@ -201,17 +220,17 @@ if [ -f "$LIVE_PREFS" ]; then
   # correctly migrated to 2.4.0 hit the else arm and FAILED as "unknown". The
   # gate was rejecting the only fully-migrated state it exists to encourage.
   #
-  # Reading the target from the migrator, and the accepted set from the schema
-  # enum, means this can never disagree with them again.
+  # Reading the target and the accepted set from the same schema enum the
+  # migrator reads means this can never disagree with it again.
   MIGRATOR="$SMOKE_DIR/migrate-prefs.mjs"
-  TARGET=$(grep -oE 'TARGET_VERSION = "[0-9.]+"' "$MIGRATOR" 2>/dev/null | grep -oE '[0-9]+\.[0-9]+\.[0-9]+')
+  TARGET=$(migration_target "$(dirname "$MIGRATOR")")
   KNOWN=$(node -e "
     const s = require('$PREFS_SCHEMA');
     process.stdout.write((s.properties.schemaVersion.enum || []).join(' '));
   " 2>/dev/null)
 
   if [ -z "$TARGET" ]; then
-    fail "cannot read TARGET_VERSION from migrate-prefs.mjs - the check has no reference point"
+    fail "cannot derive the migration target from prefs.schema.json - the check has no reference point"
   elif [ "$LVER" = "$TARGET" ]; then
     pass "live prefs at the migration target (v$TARGET)"
   elif [ "$LVER" = "none" ]; then

package/pipeline/scripts/token-budget-report.mjs

@@ -21,6 +21,7 @@
 import fs from "node:fs";
 import os from "node:os";
 import path from "node:path";
+import { rowPhase } from "../lib/phase-schema.mjs";
 
 const args = parseArgs(process.argv.slice(2));
 const METRICS =
@@ -64,7 +65,12 @@ for (const e of events) {
       models: new Set(),
     });
   const t = byTask.get(e.task_id);
-  t.phases.add(e.phase);
+  // Normalised to the current vocabulary before it goes in the set. This log
+  // is append-only across the v19.0.0 renumbering, so the raw field carries
+  // two different meanings and a set of raw values cannot be compared to
+  // anything.
+  const cp = rowPhase(e);
+  if (cp !== null) t.phases.add(cp);
   const d = e.details || {};
   if (typeof d.tokens_in === "number") t.tokens_in += d.tokens_in;
   if (typeof d.tokens_out === "number") t.tokens_out += d.tokens_out;
@@ -113,7 +119,12 @@ for (const [taskId, t] of byTask) {
   summary.totalTokensOut += t.tokens_out;
   summary.totalUsd += usd;
   const total = t.tokens_in + t.tokens_out;
-  const isMulti = t.phases.has("4") && t.premium_calls > 5; // heuristic
+  // Heuristic: a task that reached Review with many premium calls is a
+  // multi-repo run. Keyed on Review (3), not on the literal "4" this line used
+  // to carry - under the six-phase contract 4 is Commit, and every run reaches
+  // Commit, so the old literal would have classified nearly everything as
+  // multi-repo and applied the wrong ceiling.
+  const isMulti = t.phases.has(3) && t.premium_calls > 5;
   const tokenLimit = isMulti ? limits.perTaskTokensMultiRepo : limits.perTaskTokens;
   const usdLimit = isMulti ? limits.perTaskUsdMultiRepo : limits.perTaskUsd;
   if (total > tokenLimit)

package/pipeline/scripts/triage-memory.mjs

@@ -3,8 +3,8 @@
 /**
  * @file triage-memory.mjs  -  v8.3.0
  *
- * Lightweight, zero-dep persistence layer for past Phase 4 triage findings.
- * Used by Phase 7 to ingest results, by Phase 1/4 to look up prior art when
+ * Lightweight, zero-dep persistence layer for past Phase 3 triage findings.
+ * Used by Phase 5 to ingest results, by Phase 1/4 to look up prior art when
  * a new task overlaps with a previous one, and by /multi-agent:search.
  *
  * Storage: append-only JSONL, one file per repo: