@mmerterden/multi-agent-pipeline 18.0.0 → 19.1.0

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
@@ -612,8 +612,8 @@ 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.
-    # `subjects` now carries Phase 2's plan steps as indented rows, and
+    # from advancing six phases in silence.
+    # `subjects` now carries Phase 1'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
     # dependency concept; the "(bekliyor: ...)" suffix inside the step text is
@@ -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
@@ -776,9 +776,9 @@ subjects() {
     # Sub-phases ride out on the SAME list, indented in the subject string.
     #
     # The card has drawn these since sub-phases existed; the widget never has,
-    # and the widget is the surface the user actually looks at. Phase 2's plan
+    # and the widget is the surface the user actually looks at. Phase 1's plan
     # is the case that made the gap matter: the tasks, their order and their
-    # dependencies are computed, stored and used to drive Phase 3's picker, and
+    # dependencies are computed, stored and used to drive Phase 2's picker, and
     # none of it was visible anywhere the user was looking.
     #
     # Indentation is two spaces INSIDE the subject because the widget takes
@@ -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
@@ -1214,7 +1214,7 @@ GATE
     ;;
 
   plan)
-    # Phase 2's plan, turned into sub-phases of the phase that will execute it.
+    # Phase 1's plan, turned into sub-phases of the phase that will execute it.
     #
     # The parsing lives here rather than in the phase document for one reason:
     # sub-phases are this file's structure, and a jq blob in a phase doc is a
@@ -1222,7 +1222,7 @@ GATE
     # keeps the doc to one line, which the aggregate phase-doc budget cares about.
     #
     # Reads planning-output.schema.json on stdin: tasks[] with id, title and an
-    # optional dependsOn[]. Status is `pending` for all of them - Phase 3 moves
+    # optional dependsOn[]. Status is `pending` for all of them - Phase 2 moves
     # them with `sub`, and pre-marking work as started is the lie the tracker
     # exists to avoid.
     need_jq

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/pre-commit-check.sh

@@ -148,12 +148,41 @@ scan_file() {
     FOUND=1
   fi
 
-  # High-signal provider token prefixes (low false-positive rate)
+  # High-signal provider token prefixes (low false-positive rate).
+  #
+  # This set and the one in lib/outbound-gate.mjs cover the same providers, and
+  # smoke-secret-parity.sh holds them to it by running a fake token of each
+  # shape through BOTH.
   if echo "$content" | grep -qE '(ghp|gho|ghu|ghs|ghr)_[A-Za-z0-9]{36}|github_pat_[A-Za-z0-9_]{60,}|xox[baprs]-[A-Za-z0-9-]{12,}|sk_live_[A-Za-z0-9]{20,}|rk_live_[A-Za-z0-9]{20,}|AIza[0-9A-Za-z_-]{35}|npm_[A-Za-z0-9]{36}|glpat-[A-Za-z0-9_-]{20,}'; then
     echo "BLOCKED: Provider access token in $file" >&2
     FOUND=1
   fi
 
+  # Model-provider and ML-hub keys. `sk-ant-` and `sk-proj-` are checked before
+  # the bare `sk-` form so the message names the provider a reader can revoke.
+  if echo "$content" | grep -qE 'sk-ant-[A-Za-z0-9_-]{32,}|sk-proj-[A-Za-z0-9_-]{32,}|pplx-[A-Za-z0-9]{32,}|hf_[A-Za-z0-9]{30,}'; then
+    echo "BLOCKED: Model-provider API key in $file" >&2
+    FOUND=1
+  fi
+
+  # Figma personal access token (figd_) and the MCP OAuth token (figu_). Same
+  # shape, same blast radius: both read every file the account can reach.
+  if echo "$content" | grep -qE 'fig[a-z]_[A-Za-z0-9_-]{20,}'; then
+    echo "BLOCKED: Figma token in $file" >&2
+    FOUND=1
+  fi
+
+  # A URL carrying its own credentials, which is how a `git remote -v` paste
+  # leaks, and an Authorization header copied out of a curl trace.
+  if echo "$content" | grep -qE '[a-z][a-z0-9+.-]*://[^[:space:]/:@]+:[^[:space:]/@]+@'; then
+    echo "BLOCKED: URL with embedded credentials in $file" >&2
+    FOUND=1
+  fi
+  if echo "$content" | grep -qiE 'Authorization:[[:space:]]*(Bearer|Basic)[[:space:]]+[A-Za-z0-9._~+/=-]{16,}'; then
+    echo "BLOCKED: Authorization header with a token in $file" >&2
+    FOUND=1
+  fi
+
   # JWT (three base64url segments  -  header.payload.signature)
   if echo "$content" | grep -qE 'eyJ[A-Za-z0-9_-]{10,}\.eyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}'; then
     echo "BLOCKED: JWT in $file" >&2

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/scan-skills.sh

@@ -123,6 +123,14 @@ tree_grep() {
   return 0
 }
 
+# Case-insensitive variant. Prose written to steer a model is written by hand,
+# so its capitalisation is whatever the author felt like, and a case-sensitive
+# pattern would miss "Ignore all previous instructions" by one letter.
+tree_grep_i() {
+  tr '\n' '\0' < "$SCAN_LIST" | xargs -0 grep -nHEi -- "$1" 2>/dev/null
+  return 0
+}
+
 # stdin: "path:line:content" grep hits -> stdout: "fileIdx|path|line|content"
 index_hits() {
   awk -v listfile="$SCAN_LIST" '
@@ -260,6 +268,24 @@ if [ "$THRESHOLD_RANK" -ge 1 ]; then
     seq=$((seq+1))
     emit_raw "$HIT_IDX" 9 "$seq" high "$HIT_FILE" "$HIT_LINE" "chmod-then-exec" "script made executable and immediately invoked"
   done < <(tree_grep 'chmod[[:space:]]+\+x[[:space:]]+[^&;]+[[:space:]]*(&&|;)[[:space:]]*\./' | index_hits)
+
+  # Prompt injection (OWASP LLM01). The other families ask what a skill makes
+  # the MACHINE do; this one asks what it makes the MODEL do. A skill is
+  # instructions loaded straight into the context that decides everything after
+  # it, so text telling the model to drop its instructions, recite its prompt,
+  # or act behind the user's back is an attack delivered as prose - and no
+  # pattern above can see it, because nothing is executed.
+  #
+  # `high`, not `critical`: these phrasings can appear in a skill that DESCRIBES
+  # the attack, this scanner's own documentation being the obvious case, so a
+  # hit is a line a human reads rather than a verdict.
+  seq=0
+  while IFS= read -r hit; do
+    [ -z "$hit" ] && continue
+    parse_hit "$hit"
+    seq=$((seq+1))
+    emit_raw "$HIT_IDX" 13 "$seq" high "$HIT_FILE" "$HIT_LINE" "prompt-injection" "text instructing the model to override, disclose or hide - OWASP LLM01"
+  done < <(tree_grep_i 'ignore (all )?(the )?(previous|prior|above|earlier) (instructions|prompts?|rules?)|disregard (your|the|any) (system prompt|previous instructions|instructions)|(reveal|print|output|repeat|show) (your|the) (system prompt|full instructions)|(do not|don'"'"'t|never) (tell|inform) the (user|operator)|without (telling|informing|asking) the (user|operator)|(always|automatically) (approve|confirm) [^.]{0,30}(without|regardless)|you are now (a|an|the) ' | index_hits)
 fi
 
 # --- medium families (rank 2) ---------------------------------------------

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: