diff --git a/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/.openspec.yaml b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/.openspec.yaml
new file mode 100644
index 0000000..9f70866
--- /dev/null
+++ b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/.openspec.yaml
@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-05-15
diff --git a/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/proposal.md b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/proposal.md
new file mode 100644
index 0000000..937d646
--- /dev/null
+++ b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/proposal.md
@@ -0,0 +1,63 @@
+## Why
+
+`claude-supervisor.sh`'s asking/blocked classifier had two structural
+weaknesses that produced classifier cost and precision problems:
+
+1. The "is this a real interactive cursor?" gate (`last_line_is_prompt`)
+   accepted any line ending in `:` or `?` as a waiting cursor. Worker
+   tails routinely end with `Reading file:` mid-work, so an old
+   `Continue?` or `Should I` 40 lines back in the 80-line capture
+   was enough to call sonnet/medium and paste an answer codex never
+   asked for. Pure false-positive cost.
+2. `is_busy` grep'd the whole 80-line window for `Working (` or
+   `esc to interrupt`. A pane that finished `Working (12s)` 40 lines
+   ago and is now sitting at a fresh `[Y/n] ` cursor read as "busy"
+   and never reached the ask path — real asks slipped past the
+   supervisor entirely.
+
+Additionally, several real codex-fleet blockers (merge conflict,
+uncommitted-changes, fatal git, ssh permission-denied, MCP server
+missing, bad GH credentials, BLOCKED: prefix) were missing from
+`BLOCKED_PATTERNS`, so panes parked on these states classified as
+`quiet` and the supervisor stayed silent.
+
+## What Changes
+
+- Extract the classifier (BUSY/ASK/BLOCKED patterns + `is_busy`,
+  `is_asking`, `is_blocked`, `last_line_is_prompt`, `classify_tail`,
+  `tail_hash`) into a pure-bash library at
+  `scripts/codex-fleet/lib/claude-supervisor-classifier.sh` so the
+  daemon and a replay harness share one implementation.
+- Tighten `last_line_is_prompt`: drop the bare `[?:][[:space:]]*$`
+  rule. Bare-`?` is only admitted when the line carries a known
+  question lead-word. `:$` no longer counts.
+- Tighten `is_busy`: anchor BUSY_PATTERNS to the LAST non-empty line
+  only. codex rewrites the `Working (…)` footer in place; if the
+  worker is busy it's at the bottom. Stale `Working (` in scrollback
+  no longer masks a fresh interactive prompt.
+- Tighten `is_asking`: scope ASK_PATTERN matching to the recent N
+  non-empty lines (default 8 via `CLAUDE_SUPERVISOR_RECENT_LINES`)
+  AND require `last_line_is_prompt` to pass.
+- Extend `BLOCKED_PATTERNS` with the codex-fleet-specific stuck
+  states listed under "Why".
+- Add a fixture-driven replay harness at
+  `scripts/codex-fleet/test/test-claude-supervisor-classifier.sh`
+  with 24 pane-capture fixtures covering the false-positive,
+  missed-block, and previously-correct cases. Filename prefix
+  encodes the expected classification.
+
+## Impact
+
+- Cost: fewer ASK false positives → fewer sonnet/medium calls per
+  tick. Sonnet stays the workhorse for the remaining real asks.
+  Opus calls are gated on the (now more accurate) BLOCKED set;
+  strike guard caps per-pane spend.
+- Behavior: panes the supervisor used to ignore (real ask under
+  stale `Working (`, merge-conflict, MCP-missing, bad GH creds)
+  now classify correctly.
+- Risk: the harness pins the precision/recall trade-off — any
+  future loosening of the gates regresses the harness.
+- Surfaces touched: `claude-supervisor.sh` replaces its inline
+  classifier with `source`; new lib; new test + fixtures. No
+  daemon-state files, no plan-watcher changes, no cap-swap-daemon
+  changes.
diff --git a/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/specs/claude-supervisor-classifier-audit/spec.md b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/specs/claude-supervisor-classifier-audit/spec.md
new file mode 100644
index 0000000..fdbff47
--- /dev/null
+++ b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/specs/claude-supervisor-classifier-audit/spec.md
@@ -0,0 +1,89 @@
+## ADDED Requirements
+
+### Requirement: classifier lives in a pure-bash library
+
+The system SHALL keep the classifier (BUSY/ASK/BLOCKED pattern
+arrays, `is_busy`, `is_asking`, `is_blocked`, `last_line_is_prompt`,
+`classify_tail`, and `tail_hash`) at
+`scripts/codex-fleet/lib/claude-supervisor-classifier.sh`. The lib
+SHALL be safe to `source` with no side effects (no tmux calls, no
+claude calls, no file writes at source time). `claude-supervisor.sh`
+SHALL source this lib rather than inlining the classifier.
+
+#### Scenario: library is sourceable with no side effects
+- **WHEN** `scripts/codex-fleet/lib/claude-supervisor-classifier.sh`
+  is sourced from a bash script
+- **THEN** no tmux, claude, or filesystem-mutating command runs
+- **AND** `classify_tail`, `is_busy`, `is_asking`, `is_blocked`,
+  `last_line_is_prompt`, and `tail_hash` are defined
+
+### Requirement: classify_tail returns one of four labels
+
+`classify_tail "<tail>"` SHALL echo exactly one of
+`busy`, `asking`, `blocked`, `quiet`. `asking` SHALL outrank `blocked`
+when both conditions match — a pane that mentioned a stale blocker
+but is now showing an interactive menu wants an answer to the menu.
+
+#### Scenario: asking outranks blocked
+- **WHEN** the recent tail contains both a BLOCKED_PATTERN (e.g.,
+  stale-claim) and an ASK_PATTERN (e.g., a numbered menu with a
+  `(recommended)` option) AND `last_line_is_prompt` accepts the
+  bottom line
+- **THEN** `classify_tail` echoes `asking`
+
+### Requirement: busy is anchored to the last non-empty line
+
+`is_busy` SHALL match BUSY_PATTERNS only against the last non-empty
+line of the ANSI-stripped tail. A stale `Working (` or
+`esc to interrupt` earlier in scrollback SHALL NOT mask a fresh
+interactive cursor at the bottom.
+
+#### Scenario: stale Working in scrollback does not mask a fresh ask
+- **WHEN** the tail contains `Working (` nine lines from the bottom
+  AND the bottom line is a bare prompt sigil (`❯`) under a numbered
+  menu with `(recommended)` option
+- **THEN** `classify_tail` echoes `asking`, not `busy`
+
+### Requirement: last_line_is_prompt rejects bare-colon endings
+
+`last_line_is_prompt` SHALL NOT accept a bare trailing `:` as a
+waiting cursor. A bare trailing `?` SHALL be accepted only when the
+same line carries a known question lead-word (Continue, Approve,
+Proceed, Confirm, Apply, Should I, Do you want, Would you like,
+Which option/approach/one, Choose, Select, Pick, Need clarification,
+Need more …, Please clarify/confirm/choose/specify).
+
+#### Scenario: narrative status line ending in ":" does not trigger asking
+- **WHEN** the bottom line of the tail is `Reading file: …:` AND an
+  older `Continue?` appears earlier in scrollback
+- **THEN** `classify_tail` echoes `quiet`
+
+### Requirement: BLOCKED_PATTERNS cover codex-fleet stuck states
+
+`BLOCKED_PATTERNS` SHALL match each of: git merge conflict
+(`CONFLICT (content`), `error: uncommitted changes`, `fatal: ` git
+errors, `Permission denied (publickey)`, `gh: command not found`,
+`Bad credentials`, `MCP server <name> (not found|missing|unavailable)`,
+`429 Too Many Requests`, and the canonical `BLOCKED:` prefix —
+in addition to the prior set (PLAN_SUBTASK_NOT_FOUND, stale-claim,
+told-not-to-rescue, less-than-5%-limit, etc.).
+
+#### Scenario: missing MCP server is classified as blocked
+- **WHEN** the tail contains `Error: MCP server colony not found in
+  the registered servers`
+- **THEN** `classify_tail` echoes `blocked`
+
+### Requirement: replay harness pins classifier behavior
+
+The system SHALL ship a replay harness at
+`scripts/codex-fleet/test/test-claude-supervisor-classifier.sh`
+that discovers every `*.txt` fixture under
+`scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/`,
+parses the expected label from the filename prefix
+(`<label>__*.txt`), and asserts `classify_tail` returns that label.
+The harness SHALL exit non-zero if any fixture mismatches.
+
+#### Scenario: harness passes against the current classifier
+- **WHEN** the harness is run with the lib at its current head
+- **THEN** every fixture's `actual` label equals its `expected` label
+- **AND** the harness exits 0
diff --git a/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/tasks.md b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/tasks.md
new file mode 100644
index 0000000..faa7d5c
--- /dev/null
+++ b/openspec/changes/agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25/tasks.md
@@ -0,0 +1,38 @@
+## Definition of Done
+
+This change is complete only when **all** of the following are true:
+
+- Every checkbox below is checked.
+- The agent branch reaches `MERGED` state on `origin` and the PR URL + state are recorded in the completion handoff.
+- If any step blocks, append a `BLOCKED:` line under section 4 and stop.
+
+## Handoff
+
+- Handoff: change=`agent-claude-claude-supervisor-classifier-audit-2026-05-15-22-25`; branch=`agent/claude/claude-supervisor-classifier-audit-2026-05-15-22-25`; scope=`tighten the asking/blocked classifier inside claude-supervisor.sh, extract it to a pure lib, add fixture-driven replay harness`; evidence=`scripts/codex-fleet/test/test-claude-supervisor-classifier.sh`.
+
+## 1. Specification
+
+- [x] 1.1 Finalize proposal scope and acceptance criteria.
+- [x] 1.2 Define normative requirements in `specs/claude-supervisor-classifier-audit/spec.md`.
+
+## 2. Implementation
+
+- [x] 2.1 Extract classifier into `scripts/codex-fleet/lib/claude-supervisor-classifier.sh`.
+- [x] 2.2 Tighten `last_line_is_prompt` (drop bare `:$`; gate `?$` on a question lead-word).
+- [x] 2.3 Anchor `is_busy` to the last non-empty line.
+- [x] 2.4 Scope `is_asking` to recent N lines AND require `last_line_is_prompt`.
+- [x] 2.5 Extend `BLOCKED_PATTERNS` with codex-fleet stuck states.
+- [x] 2.6 Patch `claude-supervisor.sh` to `source` the new lib.
+- [x] 2.7 Add replay harness + 24 fixtures under `scripts/codex-fleet/test/`.
+
+## 3. Verification
+
+- [x] 3.1 `bash -n` clean on lib, harness, daemon.
+- [x] 3.2 `bash scripts/codex-fleet/test/test-claude-supervisor-classifier.sh` — 24 pass, 0 fail.
+- [x] 3.3 `claude-supervisor.sh --once --dry-run` runs without tmux (no-op tick, rc=0).
+
+## 4. Cleanup (mandatory; run before claiming completion)
+
+- [ ] 4.1 Run the cleanup pipeline: `gx branch finish --branch agent/claude/claude-supervisor-classifier-audit-2026-05-15-22-25 --base main --via-pr --wait-for-merge --cleanup`.
+- [ ] 4.2 Record the PR URL and final merge state (`MERGED`) in the completion handoff.
+- [ ] 4.3 Confirm the sandbox worktree is gone.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/README.md
new file mode 100644
index 0000000..0d9bc45
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/README.md
@@ -0,0 +1,30 @@
+# Plan Workspace: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+
+This folder stores durable planning artifacts before implementation changes.
+
+## Shared files
+- `summary.md`
+- `checkpoints.md`
+- `phases.md`
+- `open-questions.md`
+- `coordinator-prompt.md`
+- `kickoff-prompts.md`
+
+## Role folders
+- `planner/`
+- `architect/`
+- `critic/`
+- `executor/`
+- `writer/`
+- `verifier/`
+
+When Codex or Claude hits an unresolved question that should survive chat, add it to `open-questions.md` as an unchecked `- [ ]` item.
+
+Each role folder contains OpenSpec-style artifacts:
+- `.openspec.yaml`
+- `prompt.md` (copy/paste role prompt)
+- `proposal.md`
+- `tasks.md` (Spec / Tests / Implementation / Checkpoints checklists)
+- `specs/<role>/spec.md`
+Planner also gets `plan.md`; executor also gets `checkpoints.md`.
+Planner plans should follow `openspec/plan/PLANS.md`.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/.openspec.yaml b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/.openspec.yaml
new file mode 100644
index 0000000..ffff16e
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/.openspec.yaml
@@ -0,0 +1,9 @@
+schema: 1
+plan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+role: architect
+status: draft
+artifacts:
+  prompt: prompt.md
+  proposal: proposal.md
+  tasks: tasks.md
+  spec: specs/architect/spec.md
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/README.md
new file mode 100644
index 0000000..1e1c275
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/README.md
@@ -0,0 +1,12 @@
+# architect
+
+Role workspace for `architect`.
+
+Default artifacts:
+- `.openspec.yaml`
+- `prompt.md`
+- `proposal.md`
+- `tasks.md`
+- `specs/<role>/spec.md`
+
+Use this folder for role notes, artifacts, and status updates.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/prompt.md
new file mode 100644
index 0000000..22f9225
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/prompt.md
@@ -0,0 +1,34 @@
+# architect Prompt
+
+You are the `architect` role for OpenSpec plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Complete only this role's assigned checklist and leave compact evidence for the coordinator.
+
+## Source of truth
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/tasks.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/proposal.md`
+
+## Before edits
+
+1. Confirm branch/worktree with `git status --short --branch`.
+2. Claim every touched file before editing:
+   - Prefer Colony `task_claim_file` when an active task exists.
+   - Otherwise run `gx locks claim --branch <agent-branch> <file...>`.
+3. Stay inside assigned files/modules; coordinate before touching shared paths.
+
+## Working rules
+
+- Update `architect/tasks.md` as each item completes.
+- Record durable unresolved questions in `open-questions.md`.
+- Keep handoffs short: files changed, behavior touched, verification, risks.
+- Do not revert another agent's edits.
+
+## Cleanup
+
+Only the owner/finalizer lane runs `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`. If blocked, append `BLOCKED:` with branch, task, blocker, next, evidence.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/proposal.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/proposal.md
new file mode 100644
index 0000000..f54adf9
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/proposal.md
@@ -0,0 +1,15 @@
+# Proposal: architect (agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25)
+
+## Why
+
+Summarize why this role's work is required for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## What Changes
+
+- [ ] List the planned role-specific changes
+
+## Impact
+
+- Scope:
+- Risks:
+- Dependencies:
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/specs/architect/spec.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/specs/architect/spec.md
new file mode 100644
index 0000000..408a64d
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/specs/architect/spec.md
@@ -0,0 +1,10 @@
+# Capability Spec: architect
+
+## ADDED Requirements
+
+### Requirement: architect responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+This role MUST define and deliver its scoped outputs with evidence.
+
+#### Scenario: Role executes assigned scope
+- **WHEN** the role begins execution for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- **THEN** it follows `tasks.md` and records evidence for completion
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/tasks.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/tasks.md
new file mode 100644
index 0000000..003e055
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/architect/tasks.md
@@ -0,0 +1,33 @@
+# architect tasks
+
+## 1. Spec
+
+- [ ] 1.1 Define ownership boundaries, interfaces, and artifact responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- [ ] 1.2 Validate architecture constraints and non-functional requirements coverage
+
+## 2. Tests
+
+- [ ] 2.1 Define architectural verification checkpoints (integration boundaries, failure modes, compatibility)
+- [ ] 2.2 Validate that acceptance criteria map to concrete architecture decisions
+
+## 3. Implementation
+
+- [ ] 3.1 Review plan for strongest antithesis/tradeoff tensions
+- [ ] 3.2 Propose synthesis path and guardrails for implementation teams
+- [ ] 3.3 Record architecture sign-off notes for downstream execution
+
+## 4. Checkpoints
+
+- [ ] [A1] READY - Architecture review checkpoint
+
+## 5. Collaboration
+
+- [ ] 5.1 Owner recorded this lane before edits.
+- [ ] 5.2 Record joined agents / handoffs, or mark `N/A` when solo.
+- [ ] 5.3 Record unresolved plan questions in `../open-questions.md`, or mark `N/A` when none.
+
+## 6. Cleanup
+
+- [ ] 6.1 If this lane owns finalization, run `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`.
+- [ ] 6.2 Record PR URL + final `MERGED` state in the handoff.
+- [ ] 6.3 Confirm sandbox cleanup (`git worktree list`, `git branch -a`) or append `BLOCKED:` and stop.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md
new file mode 100644
index 0000000..ce56995
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md
@@ -0,0 +1,4 @@
+# Plan Checkpoints: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+
+Chronological checkpoint log for all roles.
+
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/coordinator-prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/coordinator-prompt.md
new file mode 100644
index 0000000..bd6a46f
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/coordinator-prompt.md
@@ -0,0 +1,41 @@
+# Master Coordinator Prompt
+
+You are the coordinator for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Drive this plan from draft to execution-ready status with strict checkpoint discipline and no scope drift.
+
+## Source-of-truth artifacts
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/plan.md`
+- role `prompt.md` files for copy/paste helper startup
+- role `tasks.md` files for planner/architect/critic/executor/writer/verifier
+
+## Coordinator responsibilities
+
+1. Keep checkpoints current in each role `tasks.md` and root `checkpoints.md`.
+2. Route unresolved questions and branching decisions into `open-questions.md`.
+3. Ensure each role has explicit acceptance criteria and verification evidence.
+4. Prevent implementation from starting before planning gates are complete.
+5. Keep handoffs concise: files changed, behavior touched, verification output, risks.
+
+## Wave-splitting decision (optional)
+
+Create wave prompts in `kickoff-prompts.md` only when at least one applies:
+
+- 3+ independent implementation lanes can run in parallel.
+- Runtime cutover/rollback sequencing needs explicit lane ownership.
+- Risk is high enough that bounded execution packets reduce coordination mistakes.
+
+If wave splitting is not needed, keep execution under a single owner with normal role checkpoints.
+
+## Exit criteria
+
+- All role checkpoints required for planning are done.
+- Execution lanes (if any) have clear ownership boundaries.
+- `open-questions.md` captures unresolved decisions that still need answers.
+- Verification plan and rollback expectations are explicit and testable.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/.openspec.yaml b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/.openspec.yaml
new file mode 100644
index 0000000..2183c61
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/.openspec.yaml
@@ -0,0 +1,9 @@
+schema: 1
+plan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+role: critic
+status: draft
+artifacts:
+  prompt: prompt.md
+  proposal: proposal.md
+  tasks: tasks.md
+  spec: specs/critic/spec.md
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/README.md
new file mode 100644
index 0000000..5b5c877
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/README.md
@@ -0,0 +1,12 @@
+# critic
+
+Role workspace for `critic`.
+
+Default artifacts:
+- `.openspec.yaml`
+- `prompt.md`
+- `proposal.md`
+- `tasks.md`
+- `specs/<role>/spec.md`
+
+Use this folder for role notes, artifacts, and status updates.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/prompt.md
new file mode 100644
index 0000000..fd55bda
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/prompt.md
@@ -0,0 +1,34 @@
+# critic Prompt
+
+You are the `critic` role for OpenSpec plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Complete only this role's assigned checklist and leave compact evidence for the coordinator.
+
+## Source of truth
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/tasks.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/proposal.md`
+
+## Before edits
+
+1. Confirm branch/worktree with `git status --short --branch`.
+2. Claim every touched file before editing:
+   - Prefer Colony `task_claim_file` when an active task exists.
+   - Otherwise run `gx locks claim --branch <agent-branch> <file...>`.
+3. Stay inside assigned files/modules; coordinate before touching shared paths.
+
+## Working rules
+
+- Update `critic/tasks.md` as each item completes.
+- Record durable unresolved questions in `open-questions.md`.
+- Keep handoffs short: files changed, behavior touched, verification, risks.
+- Do not revert another agent's edits.
+
+## Cleanup
+
+Only the owner/finalizer lane runs `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`. If blocked, append `BLOCKED:` with branch, task, blocker, next, evidence.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/proposal.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/proposal.md
new file mode 100644
index 0000000..22f729a
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/proposal.md
@@ -0,0 +1,15 @@
+# Proposal: critic (agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25)
+
+## Why
+
+Summarize why this role's work is required for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## What Changes
+
+- [ ] List the planned role-specific changes
+
+## Impact
+
+- Scope:
+- Risks:
+- Dependencies:
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/specs/critic/spec.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/specs/critic/spec.md
new file mode 100644
index 0000000..68c657f
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/specs/critic/spec.md
@@ -0,0 +1,10 @@
+# Capability Spec: critic
+
+## ADDED Requirements
+
+### Requirement: critic responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+This role MUST define and deliver its scoped outputs with evidence.
+
+#### Scenario: Role executes assigned scope
+- **WHEN** the role begins execution for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- **THEN** it follows `tasks.md` and records evidence for completion
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/tasks.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/tasks.md
new file mode 100644
index 0000000..6b455a2
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/critic/tasks.md
@@ -0,0 +1,33 @@
+# critic tasks
+
+## 1. Spec
+
+- [ ] 1.1 Validate principle-driver-option consistency across the plan
+- [ ] 1.2 Validate risks, consequences, and mitigation clarity (including idempotency expectations)
+
+## 2. Tests
+
+- [ ] 2.1 Validate testability and measurability of all acceptance criteria
+- [ ] 2.2 Validate verification steps are concrete and reproducible
+
+## 3. Implementation
+
+- [ ] 3.1 Produce verdict (APPROVE / ITERATE / REJECT) with actionable feedback
+- [ ] 3.2 Confirm revised drafts resolve prior findings before approval
+- [ ] 3.3 Publish final quality/risk sign-off notes
+
+## 4. Checkpoints
+
+- [ ] [C1] READY - Quality gate checkpoint
+
+## 5. Collaboration
+
+- [ ] 5.1 Owner recorded this lane before edits.
+- [ ] 5.2 Record joined agents / handoffs, or mark `N/A` when solo.
+- [ ] 5.3 Record unresolved plan questions in `../open-questions.md`, or mark `N/A` when none.
+
+## 6. Cleanup
+
+- [ ] 6.1 If this lane owns finalization, run `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`.
+- [ ] 6.2 Record PR URL + final `MERGED` state in the handoff.
+- [ ] 6.3 Confirm sandbox cleanup (`git worktree list`, `git branch -a`) or append `BLOCKED:` and stop.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/.openspec.yaml b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/.openspec.yaml
new file mode 100644
index 0000000..e385d7d
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/.openspec.yaml
@@ -0,0 +1,9 @@
+schema: 1
+plan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+role: executor
+status: draft
+artifacts:
+  prompt: prompt.md
+  proposal: proposal.md
+  tasks: tasks.md
+  spec: specs/executor/spec.md
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/README.md
new file mode 100644
index 0000000..62b95ee
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/README.md
@@ -0,0 +1,12 @@
+# executor
+
+Role workspace for `executor`.
+
+Default artifacts:
+- `.openspec.yaml`
+- `prompt.md`
+- `proposal.md`
+- `tasks.md`
+- `specs/<role>/spec.md`
+
+Use this folder for role notes, artifacts, and status updates.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/checkpoints.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/checkpoints.md
new file mode 100644
index 0000000..dffb6fe
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/checkpoints.md
@@ -0,0 +1,4 @@
+# executor checkpoints
+
+Timestamped execution checkpoints for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/prompt.md
new file mode 100644
index 0000000..79d785d
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/prompt.md
@@ -0,0 +1,34 @@
+# executor Prompt
+
+You are the `executor` role for OpenSpec plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Complete only this role's assigned checklist and leave compact evidence for the coordinator.
+
+## Source of truth
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/tasks.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/proposal.md`
+
+## Before edits
+
+1. Confirm branch/worktree with `git status --short --branch`.
+2. Claim every touched file before editing:
+   - Prefer Colony `task_claim_file` when an active task exists.
+   - Otherwise run `gx locks claim --branch <agent-branch> <file...>`.
+3. Stay inside assigned files/modules; coordinate before touching shared paths.
+
+## Working rules
+
+- Update `executor/tasks.md` as each item completes.
+- Record durable unresolved questions in `open-questions.md`.
+- Keep handoffs short: files changed, behavior touched, verification, risks.
+- Do not revert another agent's edits.
+
+## Cleanup
+
+Only the owner/finalizer lane runs `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`. If blocked, append `BLOCKED:` with branch, task, blocker, next, evidence.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/proposal.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/proposal.md
new file mode 100644
index 0000000..1fe425e
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/proposal.md
@@ -0,0 +1,15 @@
+# Proposal: executor (agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25)
+
+## Why
+
+Summarize why this role's work is required for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## What Changes
+
+- [ ] List the planned role-specific changes
+
+## Impact
+
+- Scope:
+- Risks:
+- Dependencies:
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/specs/executor/spec.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/specs/executor/spec.md
new file mode 100644
index 0000000..e285c96
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/specs/executor/spec.md
@@ -0,0 +1,10 @@
+# Capability Spec: executor
+
+## ADDED Requirements
+
+### Requirement: executor responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+This role MUST define and deliver its scoped outputs with evidence.
+
+#### Scenario: Role executes assigned scope
+- **WHEN** the role begins execution for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- **THEN** it follows `tasks.md` and records evidence for completion
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/tasks.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/tasks.md
new file mode 100644
index 0000000..062ff29
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/executor/tasks.md
@@ -0,0 +1,33 @@
+# executor tasks
+
+## 1. Spec
+
+- [ ] 1.1 Map approved plan requirements to concrete implementation work items
+- [ ] 1.2 Validate touched components/files are explicitly listed before coding starts
+
+## 2. Tests
+
+- [ ] 2.1 Define test additions/updates required to lock intended behavior
+- [ ] 2.2 Validate regression and smoke verification commands for delivery
+
+## 3. Implementation
+
+- [ ] 3.1 Execute implementation tasks in approved order
+- [ ] 3.2 Keep progress and evidence linked back to plan checkpoints
+- [ ] 3.3 Complete final verification bundle for handoff
+
+## 4. Checkpoints
+
+- [ ] [E1] READY - Execution start checkpoint
+
+## 5. Collaboration
+
+- [ ] 5.1 Owner recorded this lane before edits.
+- [ ] 5.2 Record joined agents / handoffs, or mark `N/A` when solo.
+- [ ] 5.3 Record unresolved plan questions in `../open-questions.md`, or mark `N/A` when none.
+
+## 6. Cleanup
+
+- [ ] 6.1 If this lane owns finalization, run `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`.
+- [ ] 6.2 Record PR URL + final `MERGED` state in the handoff.
+- [ ] 6.3 Confirm sandbox cleanup (`git worktree list`, `git branch -a`) or append `BLOCKED:` and stop.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/kickoff-prompts.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/kickoff-prompts.md
new file mode 100644
index 0000000..3fbd32d
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/kickoff-prompts.md
@@ -0,0 +1,108 @@
+# Kickoff Prompts (Copy/Paste)
+
+Use these only when the coordinator decides wave-splitting is needed.
+
+## Prompt A — Wave A (Primary lane)
+
+```text
+You own Wave-A for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25` in /home/deadpool/Documents/codex-lb.
+
+Goal:
+Implement the assigned Wave-A scope and return verification evidence.
+
+Hard constraints:
+- You are not alone in the codebase; do not revert others' work.
+- Stay in your owned files/modules only.
+- Record explicit handoff notes for integration.
+
+Owned scope:
+- <fill owned files/modules>
+
+Verification:
+- <fill commands>
+
+Handoff format:
+- Files changed
+- Behavior touched
+- Verification outputs
+- Risks/follow-ups
+```
+
+## Prompt B — Wave B (Secondary lane)
+
+```text
+You own Wave-B for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25` in /home/deadpool/Documents/codex-lb.
+
+Goal:
+Implement the assigned Wave-B scope and return verification evidence.
+
+Hard constraints:
+- You are not alone in the codebase; do not revert others' work.
+- Stay in your owned files/modules only.
+- Record explicit handoff notes for integration.
+
+Owned scope:
+- <fill owned files/modules>
+
+Verification:
+- <fill commands>
+
+Handoff format:
+- Files changed
+- Behavior touched
+- Verification outputs
+- Risks/follow-ups
+```
+
+## Prompt C — Wave C (Secondary lane)
+
+```text
+You own Wave-C for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25` in /home/deadpool/Documents/codex-lb.
+
+Goal:
+Implement the assigned Wave-C scope and return verification evidence.
+
+Hard constraints:
+- You are not alone in the codebase; do not revert others' work.
+- Stay in your owned files/modules only.
+- Record explicit handoff notes for integration.
+
+Owned scope:
+- <fill owned files/modules>
+
+Verification:
+- <fill commands>
+
+Handoff format:
+- Files changed
+- Behavior touched
+- Verification outputs
+- Risks/follow-ups
+```
+
+## Prompt D — Integrator lane
+
+```text
+You are the integrator for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25` in /home/deadpool/Documents/codex-lb.
+
+Goal:
+Integrate completed waves, resolve conflicts, run final verification, and prepare rollout/cutover notes.
+
+Hard constraints:
+- You are not alone in the codebase; do not revert others' work.
+- Preserve safety-critical behavior unless explicitly planned and tested.
+- Keep final output evidence-first.
+
+Owned scope:
+- integration glue and shared touchpoints
+- final validation + handoff summary
+
+Verification:
+- <fill commands>
+
+Final report:
+- Files changed
+- Integration decisions
+- Verification outputs
+- Remaining risks
+```
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md
new file mode 100644
index 0000000..9c19bc3
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md
@@ -0,0 +1,6 @@
+# Open Questions: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+
+Capture unresolved plan questions here as unchecked checklist items.
+Keep each item concrete, decision-shaped, and easy to close with evidence.
+
+- [ ] Add the next unresolved question here.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/phases.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/phases.md
new file mode 100644
index 0000000..437f4c7
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/phases.md
@@ -0,0 +1,15 @@
+# Plan Phases: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+
+One entry per phase. Checkbox marks map to: `x` = completed, `>` = in progress, space = pending.
+Indented sub-bullets are optional metadata consumed by the Plans UI:
+
+- `session`: which agent kind runs the phase (`codex` / `claude`).
+- `checkpoints`: comma-separated role checkpoint ids delivered within the phase.
+- `summary`: one short sentence rendered under the phase title.
+
+One phase is intended to fit into a single Codex or Claude session task.
+
+- [ ] [PH01] First milestone title goes here
+  - session: codex
+  - checkpoints: P1, A1
+  - summary: Describe the single session outcome expected for this phase.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/.openspec.yaml b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/.openspec.yaml
new file mode 100644
index 0000000..520f7e2
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/.openspec.yaml
@@ -0,0 +1,9 @@
+schema: 1
+plan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+role: planner
+status: draft
+artifacts:
+  prompt: prompt.md
+  proposal: proposal.md
+  tasks: tasks.md
+  spec: specs/planner/spec.md
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/README.md
new file mode 100644
index 0000000..1b0ad5d
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/README.md
@@ -0,0 +1,12 @@
+# planner
+
+Role workspace for `planner`.
+
+Default artifacts:
+- `.openspec.yaml`
+- `prompt.md`
+- `proposal.md`
+- `tasks.md`
+- `specs/<role>/spec.md`
+
+Use this folder for role notes, artifacts, and status updates.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/plan.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/plan.md
new file mode 100644
index 0000000..5f6a160
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/plan.md
@@ -0,0 +1,65 @@
+# ExecPlan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+
+This ExecPlan is a living document. Keep `Progress`, `Surprises & Discoveries`, `Decision Log`, and `Outcomes & Retrospective` current as work proceeds.
+
+Follow repository guidance in `openspec/plan/PLANS.md`.
+
+## Purpose / Big Picture
+
+Describe what becomes possible after this plan is executed and how a user/operator can observe it working.
+
+## Progress
+
+- [ ] (YYYY-MM-DD HH:MMZ) Capture initial scope and acceptance criteria.
+- [ ] (YYYY-MM-DD HH:MMZ) Draft architecture/tradeoff plan and verification strategy.
+- [ ] (YYYY-MM-DD HH:MMZ) Finalize execution-ready handoff.
+
+## Surprises & Discoveries
+
+- Observation: _none yet_
+  Evidence: _n/a_
+
+## Decision Log
+
+- Decision: Use OpenSpec plan workspace as source of truth for this planning cycle.
+  Rationale: Keeps planning artifacts in-repo and reviewable.
+  Date/Author: YYYY-MM-DD / planner
+
+## Outcomes & Retrospective
+
+Summarize outcomes, gaps, and lessons learned when a milestone or the full plan is completed.
+
+## Context and Orientation
+
+Describe relevant modules, files, constraints, and assumptions for a newcomer. Use repository-relative paths.
+
+## Plan of Work
+
+Describe the sequence of edits and deliverables in prose. Name target files and expected effects.
+
+## Concrete Steps
+
+List exact commands with working directory and short expected outcomes.
+
+    cd /home/deadpool/Documents/codex-lb
+    openspec validate --specs
+
+## Validation and Acceptance
+
+State observable behavior and verification evidence required before execution handoff.
+
+## Idempotence and Recovery
+
+Document safe re-run behavior, rollback strategy, and failure recovery notes.
+
+## Artifacts and Notes
+
+Capture concise command output snippets, evidence pointers, and references.
+
+## Interfaces and Dependencies
+
+Name concrete interfaces/modules/dependencies and any required signatures/contracts.
+
+## Revision Note
+
+- YYYY-MM-DD HH:MMZ: Initial scaffold generated by `scripts/openspec/init-plan-workspace.sh`.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/prompt.md
new file mode 100644
index 0000000..4c66aa2
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/prompt.md
@@ -0,0 +1,34 @@
+# planner Prompt
+
+You are the `planner` role for OpenSpec plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Complete only this role's assigned checklist and leave compact evidence for the coordinator.
+
+## Source of truth
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/tasks.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/proposal.md`
+
+## Before edits
+
+1. Confirm branch/worktree with `git status --short --branch`.
+2. Claim every touched file before editing:
+   - Prefer Colony `task_claim_file` when an active task exists.
+   - Otherwise run `gx locks claim --branch <agent-branch> <file...>`.
+3. Stay inside assigned files/modules; coordinate before touching shared paths.
+
+## Working rules
+
+- Update `planner/tasks.md` as each item completes.
+- Record durable unresolved questions in `open-questions.md`.
+- Keep handoffs short: files changed, behavior touched, verification, risks.
+- Do not revert another agent's edits.
+
+## Cleanup
+
+Only the owner/finalizer lane runs `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`. If blocked, append `BLOCKED:` with branch, task, blocker, next, evidence.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/proposal.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/proposal.md
new file mode 100644
index 0000000..702789c
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/proposal.md
@@ -0,0 +1,15 @@
+# Proposal: planner (agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25)
+
+## Why
+
+Summarize why this role's work is required for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## What Changes
+
+- [ ] List the planned role-specific changes
+
+## Impact
+
+- Scope:
+- Risks:
+- Dependencies:
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/specs/planner/spec.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/specs/planner/spec.md
new file mode 100644
index 0000000..44295fe
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/specs/planner/spec.md
@@ -0,0 +1,10 @@
+# Capability Spec: planner
+
+## ADDED Requirements
+
+### Requirement: planner responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+This role MUST define and deliver its scoped outputs with evidence.
+
+#### Scenario: Role executes assigned scope
+- **WHEN** the role begins execution for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- **THEN** it follows `tasks.md` and records evidence for completion
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/tasks.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/tasks.md
new file mode 100644
index 0000000..1ea954b
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/planner/tasks.md
@@ -0,0 +1,33 @@
+# planner tasks
+
+## 1. Spec
+
+- [ ] 1.1 Define planning principles, decision drivers, and viable options for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- [ ] 1.2 Validate that scope, constraints, and acceptance criteria are captured in `summary.md`
+
+## 2. Tests
+
+- [ ] 2.1 Define verification approach for plan quality (traceability, testability, evidence expectations)
+- [ ] 2.2 Validate OpenSpec consistency checkpoints (including `openspec validate --specs` when applicable)
+
+## 3. Implementation
+
+- [ ] 3.1 Produce the initial RALPLAN-DR plan draft
+- [ ] 3.2 Integrate Architect/Critic feedback into revised plan iterations
+- [ ] 3.3 Publish final planning handoff with explicit execution lanes
+
+## 4. Checkpoints
+
+- [ ] [P1] READY - Initial planning draft checkpoint
+
+## 5. Collaboration
+
+- [ ] 5.1 Owner recorded this lane before edits.
+- [ ] 5.2 Record joined agents / handoffs, or mark `N/A` when solo.
+- [ ] 5.3 Record unresolved plan questions in `../open-questions.md`, or mark `N/A` when none.
+
+## 6. Cleanup
+
+- [ ] 6.1 If this lane owns finalization, run `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`.
+- [ ] 6.2 Record PR URL + final `MERGED` state in the handoff.
+- [ ] 6.3 Confirm sandbox cleanup (`git worktree list`, `git branch -a`) or append `BLOCKED:` and stop.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md
new file mode 100644
index 0000000..5edb6fa
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md
@@ -0,0 +1,8 @@
+# Plan Summary: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+
+- **Mode:** ralplan
+- **Status:** draft
+
+## Context
+
+Describe the planning context, constraints, and desired outcomes.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/.openspec.yaml b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/.openspec.yaml
new file mode 100644
index 0000000..ea9a262
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/.openspec.yaml
@@ -0,0 +1,9 @@
+schema: 1
+plan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+role: verifier
+status: draft
+artifacts:
+  prompt: prompt.md
+  proposal: proposal.md
+  tasks: tasks.md
+  spec: specs/verifier/spec.md
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/README.md
new file mode 100644
index 0000000..1daa373
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/README.md
@@ -0,0 +1,12 @@
+# verifier
+
+Role workspace for `verifier`.
+
+Default artifacts:
+- `.openspec.yaml`
+- `prompt.md`
+- `proposal.md`
+- `tasks.md`
+- `specs/<role>/spec.md`
+
+Use this folder for role notes, artifacts, and status updates.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/prompt.md
new file mode 100644
index 0000000..969d982
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/prompt.md
@@ -0,0 +1,34 @@
+# verifier Prompt
+
+You are the `verifier` role for OpenSpec plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Complete only this role's assigned checklist and leave compact evidence for the coordinator.
+
+## Source of truth
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/tasks.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/proposal.md`
+
+## Before edits
+
+1. Confirm branch/worktree with `git status --short --branch`.
+2. Claim every touched file before editing:
+   - Prefer Colony `task_claim_file` when an active task exists.
+   - Otherwise run `gx locks claim --branch <agent-branch> <file...>`.
+3. Stay inside assigned files/modules; coordinate before touching shared paths.
+
+## Working rules
+
+- Update `verifier/tasks.md` as each item completes.
+- Record durable unresolved questions in `open-questions.md`.
+- Keep handoffs short: files changed, behavior touched, verification, risks.
+- Do not revert another agent's edits.
+
+## Cleanup
+
+Only the owner/finalizer lane runs `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`. If blocked, append `BLOCKED:` with branch, task, blocker, next, evidence.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/proposal.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/proposal.md
new file mode 100644
index 0000000..c54f137
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/proposal.md
@@ -0,0 +1,15 @@
+# Proposal: verifier (agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25)
+
+## Why
+
+Summarize why this role's work is required for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## What Changes
+
+- [ ] List the planned role-specific changes
+
+## Impact
+
+- Scope:
+- Risks:
+- Dependencies:
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/specs/verifier/spec.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/specs/verifier/spec.md
new file mode 100644
index 0000000..cad899d
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/specs/verifier/spec.md
@@ -0,0 +1,10 @@
+# Capability Spec: verifier
+
+## ADDED Requirements
+
+### Requirement: verifier responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+This role MUST define and deliver its scoped outputs with evidence.
+
+#### Scenario: Role executes assigned scope
+- **WHEN** the role begins execution for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- **THEN** it follows `tasks.md` and records evidence for completion
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/tasks.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/tasks.md
new file mode 100644
index 0000000..6c4e9c4
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/verifier/tasks.md
@@ -0,0 +1,33 @@
+# verifier tasks
+
+## 1. Spec
+
+- [ ] 1.1 Define end-to-end validation matrix for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- [ ] 1.2 Validate success/failure conditions and evidence requirements
+
+## 2. Tests
+
+- [ ] 2.1 Execute verification commands and collect outputs
+- [ ] 2.2 Validate idempotency/re-run behavior and error-path handling
+
+## 3. Implementation
+
+- [ ] 3.1 Verify completed work against acceptance criteria
+- [ ] 3.2 Produce pass/fail findings with concrete evidence links
+- [ ] 3.3 Publish final verification sign-off (or blocker report)
+
+## 4. Checkpoints
+
+- [ ] [V1] READY - Verification checkpoint
+
+## 5. Collaboration
+
+- [ ] 5.1 Owner recorded this lane before edits.
+- [ ] 5.2 Record joined agents / handoffs, or mark `N/A` when solo.
+- [ ] 5.3 Record unresolved plan questions in `../open-questions.md`, or mark `N/A` when none.
+
+## 6. Cleanup
+
+- [ ] 6.1 If this lane owns finalization, run `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`.
+- [ ] 6.2 Record PR URL + final `MERGED` state in the handoff.
+- [ ] 6.3 Confirm sandbox cleanup (`git worktree list`, `git branch -a`) or append `BLOCKED:` and stop.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/.openspec.yaml b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/.openspec.yaml
new file mode 100644
index 0000000..b653b7b
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/.openspec.yaml
@@ -0,0 +1,9 @@
+schema: 1
+plan: agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25
+role: writer
+status: draft
+artifacts:
+  prompt: prompt.md
+  proposal: proposal.md
+  tasks: tasks.md
+  spec: specs/writer/spec.md
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/README.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/README.md
new file mode 100644
index 0000000..757c824
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/README.md
@@ -0,0 +1,12 @@
+# writer
+
+Role workspace for `writer`.
+
+Default artifacts:
+- `.openspec.yaml`
+- `prompt.md`
+- `proposal.md`
+- `tasks.md`
+- `specs/<role>/spec.md`
+
+Use this folder for role notes, artifacts, and status updates.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/prompt.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/prompt.md
new file mode 100644
index 0000000..638ce0e
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/prompt.md
@@ -0,0 +1,34 @@
+# writer Prompt
+
+You are the `writer` role for OpenSpec plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## Objective
+
+Complete only this role's assigned checklist and leave compact evidence for the coordinator.
+
+## Source of truth
+
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/summary.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/checkpoints.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/open-questions.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/tasks.md`
+- `openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/proposal.md`
+
+## Before edits
+
+1. Confirm branch/worktree with `git status --short --branch`.
+2. Claim every touched file before editing:
+   - Prefer Colony `task_claim_file` when an active task exists.
+   - Otherwise run `gx locks claim --branch <agent-branch> <file...>`.
+3. Stay inside assigned files/modules; coordinate before touching shared paths.
+
+## Working rules
+
+- Update `writer/tasks.md` as each item completes.
+- Record durable unresolved questions in `open-questions.md`.
+- Keep handoffs short: files changed, behavior touched, verification, risks.
+- Do not revert another agent's edits.
+
+## Cleanup
+
+Only the owner/finalizer lane runs `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`. If blocked, append `BLOCKED:` with branch, task, blocker, next, evidence.
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/proposal.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/proposal.md
new file mode 100644
index 0000000..d744c42
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/proposal.md
@@ -0,0 +1,15 @@
+# Proposal: writer (agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25)
+
+## Why
+
+Summarize why this role's work is required for plan `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`.
+
+## What Changes
+
+- [ ] List the planned role-specific changes
+
+## Impact
+
+- Scope:
+- Risks:
+- Dependencies:
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/specs/writer/spec.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/specs/writer/spec.md
new file mode 100644
index 0000000..d663835
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/specs/writer/spec.md
@@ -0,0 +1,10 @@
+# Capability Spec: writer
+
+## ADDED Requirements
+
+### Requirement: writer responsibilities for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+This role MUST define and deliver its scoped outputs with evidence.
+
+#### Scenario: Role executes assigned scope
+- **WHEN** the role begins execution for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- **THEN** it follows `tasks.md` and records evidence for completion
diff --git a/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/tasks.md b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/tasks.md
new file mode 100644
index 0000000..6980597
--- /dev/null
+++ b/openspec/plan/agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25/writer/tasks.md
@@ -0,0 +1,33 @@
+# writer tasks
+
+## 1. Spec
+
+- [ ] 1.1 Validate documentation scope and audience for `agent-claude-masterplan-claude-supervisor-classifier-audit-2026-05-15-22-25`
+- [ ] 1.2 Validate consistency between plan terminology and OpenSpec artifacts
+
+## 2. Tests
+
+- [ ] 2.1 Define documentation verification checklist (accuracy, completeness, command correctness)
+- [ ] 2.2 Validate command/help text examples against current workflow behavior
+
+## 3. Implementation
+
+- [ ] 3.1 Update workflow docs and command guidance for approved plan behavior
+- [ ] 3.2 Add or refine examples for operator usage and handoff clarity
+- [ ] 3.3 Publish final docs change summary with references
+
+## 4. Checkpoints
+
+- [ ] [W1] READY - Docs update checkpoint
+
+## 5. Collaboration
+
+- [ ] 5.1 Owner recorded this lane before edits.
+- [ ] 5.2 Record joined agents / handoffs, or mark `N/A` when solo.
+- [ ] 5.3 Record unresolved plan questions in `../open-questions.md`, or mark `N/A` when none.
+
+## 6. Cleanup
+
+- [ ] 6.1 If this lane owns finalization, run `gx branch finish --branch <agent-branch> --base dev --via-pr --wait-for-merge --cleanup`.
+- [ ] 6.2 Record PR URL + final `MERGED` state in the handoff.
+- [ ] 6.3 Confirm sandbox cleanup (`git worktree list`, `git branch -a`) or append `BLOCKED:` and stop.
diff --git a/scripts/codex-fleet/claude-supervisor.sh b/scripts/codex-fleet/claude-supervisor.sh
new file mode 100755
index 0000000..fd9f92e
--- /dev/null
+++ b/scripts/codex-fleet/claude-supervisor.sh
@@ -0,0 +1,565 @@
+#!/usr/bin/env bash
+# claude-supervisor.sh — two-way claude-code supervisor for codex-fleet
+# panes. Replaces the dumb "paste the same OVERRIDE forever" loop in
+# plan-watcher.sh with a read-think-respond cycle: the supervisor reads
+# what the worker said (capture-pane), feeds that to `claude -p` along
+# with the active plan + Colony queue state, and pastes claude's tailored
+# response back into the pane.
+#
+# Why this exists:
+#   The operator caught plan-watcher.sh blasting an idle worker with
+#   "OVERRIDE: claim sub-task 0 of plan X" while the worker was already
+#   posting "Blocked. PLAN_SUBTASK_NOT_FOUND, queue empty, stale-claim
+#   blocker, told me not to rescue, what now?". Pure monologue — no
+#   collaboration. This daemon flips that to dialog: claude reads the
+#   blocker, decides whether to rescue / switch plan / wait, and pastes
+#   the right prompt back.
+#
+# Architecture (one tick):
+#   1. tmux list-panes -> worker panes (skip header + non-codex panes)
+#   2. For each pane:
+#        a. capture-pane -p -S -80 -> last 80 lines
+#        b. classify: working | idle-polling | blocked | stale-blocked
+#        c. if state >= blocked AND outside cooldown:
+#             - build context prompt (pane tail + active plan + queue snapshot)
+#             - claude -p --output-format text < prompt
+#             - parse ACTION:/PROMPT:/TOOL: from response
+#             - apply: optional MCP/cli call, then paste-buffer + Enter
+#        d. record (pane, response-hash) so we don't re-ask claude with
+#           the same input ("worker said the same thing, supervisor already
+#           answered, paste-once + cooldown")
+#
+# What this does NOT do:
+#   - Run `claude -p` for every idle pane every tick. That's $50/hour.
+#     It only fires when (a) the worker emitted a real blocker / stale
+#     pattern, AND (b) we haven't asked claude this exact question yet,
+#     AND (c) the per-pane cooldown has elapsed.
+#   - Replace plan-watcher.sh. plan-watcher is the fast path for "idle
+#     worker, plan with available sub-tasks" (claude doesn't need to
+#     think — just paste the next one). claude-supervisor is the slow
+#     path for "worker hit a wall, what now". Both run in parallel.
+#   - Bypass operator approvals. tmux send-keys goes through the same
+#     pane-write contract as the rest of the fleet — capture first,
+#     refuse `bash` panes, dedup by content hash.
+#
+# Env:
+#   CODEX_FLEET_SESSION             (default codex-fleet)
+#   CODEX_FLEET_REPO_ROOT           (default git toplevel)
+#   CLAUDE_SUPERVISOR_INTERVAL      (default 60s)
+#   CLAUDE_SUPERVISOR_COOLDOWN      (default 300s — per-pane min gap between claude calls)
+#   CLAUDE_SUPERVISOR_MIN_BLOCKED   (default 120s — pane must be blocked this long before asking claude)
+#   CLAUDE_BIN                      (default `claude`)
+#   CLAUDE_MODEL_ASKING             (default `sonnet`        — fast/cheap menu picks)
+#   CLAUDE_EFFORT_ASKING            (default `medium`        — picking option 3 doesn't need xhigh)
+#   CLAUDE_MODEL_BLOCKED            (default `claude-opus-4-7` — real reasoning for genuinely stuck panes)
+#   CLAUDE_EFFORT_BLOCKED           (default `high`)
+#   CLAUDE_MODEL                    (legacy alias for CLAUDE_MODEL_BLOCKED — kept for back-compat)
+#   CLAUDE_EFFORT                   (legacy alias for CLAUDE_EFFORT_BLOCKED)
+#   CLAUDE_FALLBACK_MODEL           (default `sonnet` — when the primary is overloaded)
+#   CLAUDE_SUPERVISOR_STATE_DIR     (default $FLEET_STATE_DIR/claude-supervisor)
+#   CLAUDE_SUPERVISOR_STRIKE_WINDOW (default 3600s — sliding window for 3-strike loop guard)
+#   CLAUDE_SUPERVISOR_STRIKE_LIMIT  (default 3      — escalate after this many pastes inside the window)
+#
+# Usage:
+#   bash claude-supervisor.sh --loop                # default 60s cadence
+#   bash claude-supervisor.sh --once                # one tick + exit
+#   bash claude-supervisor.sh --once --dry-run      # show decisions, don't send
+#   bash claude-supervisor.sh --once --pane %298    # single pane
+
+set -eo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="${CODEX_FLEET_REPO_ROOT:-$(cd "$SCRIPT_DIR/../.." && pwd)}"
+SESSION="${CODEX_FLEET_SESSION:-codex-fleet}"
+WINDOW="${CODEX_FLEET_OVERVIEW_WINDOW:-overview}"
+FLEET_STATE_DIR="${FLEET_STATE_DIR:-/tmp/claude-viz}"
+STATE_DIR="${CLAUDE_SUPERVISOR_STATE_DIR:-$FLEET_STATE_DIR/claude-supervisor}"
+INTERVAL="${CLAUDE_SUPERVISOR_INTERVAL:-60}"
+COOLDOWN="${CLAUDE_SUPERVISOR_COOLDOWN:-300}"
+MIN_BLOCKED="${CLAUDE_SUPERVISOR_MIN_BLOCKED:-120}"
+CLAUDE_BIN="${CLAUDE_BIN:-claude}"
+# Back-compat: CLAUDE_MODEL / CLAUDE_EFFORT used to drive every call. They now
+# default the BLOCKED tier (the path that genuinely needs reasoning).
+CLAUDE_MODEL_BLOCKED="${CLAUDE_MODEL_BLOCKED:-${CLAUDE_MODEL:-claude-opus-4-7}}"
+CLAUDE_EFFORT_BLOCKED="${CLAUDE_EFFORT_BLOCKED:-${CLAUDE_EFFORT:-high}}"
+CLAUDE_MODEL_ASKING="${CLAUDE_MODEL_ASKING:-sonnet}"
+CLAUDE_EFFORT_ASKING="${CLAUDE_EFFORT_ASKING:-medium}"
+CLAUDE_FALLBACK_MODEL="${CLAUDE_FALLBACK_MODEL:-sonnet}"
+STRIKE_WINDOW="${CLAUDE_SUPERVISOR_STRIKE_WINDOW:-3600}"
+STRIKE_LIMIT="${CLAUDE_SUPERVISOR_STRIKE_LIMIT:-3}"
+
+# JSON Schema enforced via `claude --json-schema`. Keeps `apply_response`'s
+# parser from having to guess what claude wrote — the CLI validates against
+# this schema before returning, so we either get well-formed JSON or a
+# non-zero exit (which falls through to the legacy text parser).
+JSON_SCHEMA='{
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["action", "plan", "prompt"],
+  "properties": {
+    "action": {"type": "string", "enum": ["A", "B", "C", "D"]},
+    "plan":   {"type": "string"},
+    "tool":   {"type": "string"},
+    "prompt": {"type": "string"}
+  }
+}'
+
+# Build the claude args for a given supervisor state. Each flag is
+# conditional so an empty env var wipes that one flag instead of injecting
+# `--model ""`. `-p` and `--json-schema` come from the caller's invocation.
+claude_args_for() {
+  local state="$1"
+  local model effort
+  case "$state" in
+    asking)  model="$CLAUDE_MODEL_ASKING";  effort="$CLAUDE_EFFORT_ASKING"  ;;
+    *)       model="$CLAUDE_MODEL_BLOCKED"; effort="$CLAUDE_EFFORT_BLOCKED" ;;
+  esac
+  local -a args=( --output-format text )
+  [ -n "$model" ]                 && args+=( --model "$model" )
+  [ -n "$effort" ]                && args+=( --effort "$effort" )
+  [ -n "$CLAUDE_FALLBACK_MODEL" ] && args+=( --fallback-model "$CLAUDE_FALLBACK_MODEL" )
+  args+=( --json-schema "$JSON_SCHEMA" )
+  # Prompt cache: append our supervisor role/rules from a stable file
+  # AND exclude per-machine sections (cwd, env, git status, time) from
+  # the default system prompt so the prefix hashes identically across
+  # calls. ~50-80% Opus discount on repeat ticks.
+  if [ -f "$SYSTEM_PROMPT_FILE" ]; then
+    args+=( --append-system-prompt-file "$SYSTEM_PROMPT_FILE" )
+    args+=( --exclude-dynamic-system-prompt-sections )
+  fi
+  printf '%s\n' "${args[@]}"
+}
+
+ONCE=0
+DRY_RUN=0
+ONLY_PANE=""
+
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --loop)     ONCE=0; shift ;;
+    --once)     ONCE=1; shift ;;
+    --dry-run)  DRY_RUN=1; shift ;;
+    --pane)     ONLY_PANE="$2"; shift 2 ;;
+    --interval) INTERVAL="$2"; shift 2 ;;
+    -h|--help)  sed -n '1,40p' "$0"; exit 0 ;;
+    *) echo "claude-supervisor: unknown flag $1" >&2; exit 2 ;;
+  esac
+done
+
+mkdir -p "$STATE_DIR"
+LOG="$STATE_DIR/log"
+ts() { date +%H:%M:%S; }
+log() { printf '[%s] claude-supervisor: %s\n' "$(ts)" "$*" | tee -a "$LOG"; }
+
+# ── prompt-cache helpers ───────────────────────────────────────────────────
+# The static role + decision rules + output-format spec are byte-identical
+# across calls, so we put them in an append-system-prompt file. With
+# --exclude-dynamic-system-prompt-sections, Anthropic's prompt cache hits
+# on the full system-prompt prefix. The per-pane variable bits (plan
+# content, tail, prior answers) live in the user message and don't get
+# cached, which is fine — they're small.
+SYSTEM_PROMPT_FILE="$STATE_DIR/system-prompt.txt"
+write_system_prompt() {
+  cat > "$SYSTEM_PROMPT_FILE" <<'SYSPROMPT'
+You are the supervisor of a codex-fleet worker pane. The operator just
+captured the worker's recent scrollback and pinned the active plan; you
+decide one next action. Your reply is pasted verbatim into the worker's
+codex CLI, so write it as the operator would type it — no quotation
+marks, no "tell the worker to" prefix.
+
+# State semantics
+- state=asking — codex is showing a waiting-on-input prompt: a numbered
+  menu, "(recommended)"/"(default)" tag, [Y/n], "Continue?", etc. Pick
+  one and move on; don't lecture.
+- state=blocked — the worker emitted a genuine blocker pattern
+  (PLAN_SUBTASK_NOT_FOUND, stale-claim, "less than 5% of your 5h
+  limit", etc.). Decide rescue / switch-plan / wait.
+
+# Decision rules — state=asking
+- If a literal "(recommended)" or "(default)" tag appears on an
+  option, pick THAT option. Only override when the recommended option
+  would clearly break an acceptance criterion of the active plan.
+- If no option is tagged, prefer the lowest-risk reversible choice:
+  read-only > local edit > new branch > merge > destructive op. When
+  in doubt, the option that keeps the active plan's branch/worktree
+  intact.
+- For yes/no prompts (Y/n, y/N, Continue?, Approve?, Proceed?):
+  answer Y when the worker is mid-implementation on a task it
+  legitimately claimed and the action is in-scope; answer n when the
+  action is destructive outside its claim.
+- Keep PROMPT short. A menu pick is one option number or letter.
+
+# Decision rules — state=blocked
+- "stale-claim blocker, you told me not to rescue" → either authorize
+  the rescue with the exact mcp__colony__rescue_stranded_run call, OR
+  abandon that lane and switch to task_ready_for_agent against another
+  plan with available > 0.
+- "queue is empty" but the plan registry shows available > 0 on a
+  plan → name that plan + sub-task explicitly.
+- "less than 5% of your 5h limit" → tell worker to post a brief
+  blocker, hand off, and exit (cap-swap-daemon respawns).
+- Worker is mid-task / healthy → action=C / prompt=(none).
+- Never blast the same OVERRIDE the worker already complained about.
+- Never claim a sub-task the registry shows as completed.
+- Keep PROMPT under 600 characters.
+
+# Reply format — JSON, schema-enforced by the CLI
+{
+  "action": "A" | "B" | "C" | "D",
+  "plan":   "one short sentence: what you're about to make the worker do and why; '(none)' if action=C",
+  "tool":   "optional MCP tool name + minimal args, or '(none)'",
+  "prompt": "text to paste verbatim into the worker pane, or '(none)' if action=C"
+}
+A=rescue stale claim, B=switch plan, C=do nothing, D=answer codex question.
+Emit the JSON object and nothing else. No markdown, no preamble.
+SYSPROMPT
+}
+write_system_prompt
+
+# Classifier (BUSY/ASK/BLOCKED patterns + is_busy / is_asking / is_blocked /
+# last_line_is_prompt / classify_tail / tail_hash) lives in a pure-bash lib
+# so the replay harness under scripts/codex-fleet/test/ can exercise it
+# without spinning up the daemon. Sourcing has no side effects.
+. "$SCRIPT_DIR/lib/claude-supervisor-classifier.sh"
+
+# Per-pane state TSV: pane_id <TAB> last_ts <TAB> last_hash
+STATE_FILE="$STATE_DIR/panes.tsv"
+[ -f "$STATE_FILE" ] || : > "$STATE_FILE"
+
+# Paste log: ts <TAB> pane_id <TAB> state. One row per supervisor paste,
+# used by the 3-strike guard to detect "claude keeps answering and codex
+# keeps re-asking" loops. Survives restarts; the strike helper only counts
+# rows inside the sliding window so old entries decay automatically.
+PASTE_LOG="$STATE_DIR/paste-log.tsv"
+[ -f "$PASTE_LOG" ] || : > "$PASTE_LOG"
+
+# Operator escalation feed: ts <TAB> pane_id <TAB> state <TAB> reason.
+# When a pane crosses STRIKE_LIMIT we write here and skip the call; the
+# operator (or another tool) reads this file to notice loops the
+# supervisor refused to keep paying for.
+ESCALATE_LOG="$STATE_DIR/escalate.tsv"
+[ -f "$ESCALATE_LOG" ] || : > "$ESCALATE_LOG"
+
+# Last-response per pane: pane_id <TAB> ts <TAB> state <TAB> action <TAB> prompt(escaped).
+# Surfaces "you already answered this pane with X" to the next claude
+# call so multi-step dialogs stay coherent ("I already picked option 3,
+# this follow-up menu is the next step").
+LAST_RESPONSE_FILE="$STATE_DIR/last-response.tsv"
+[ -f "$LAST_RESPONSE_FILE" ] || : > "$LAST_RESPONSE_FILE"
+
+record_last_response() {
+  local pane="$1" state="$2" action="$3" prompt="$4" now; now="$(date +%s)"
+  # Newlines/tabs in prompt would break TSV; collapse to single-space.
+  local clean_prompt
+  clean_prompt="$(printf '%s' "$prompt" | tr '\t\n' '  ' | head -c 400)"
+  local tmp; tmp="$(mktemp)"
+  awk -F'\t' -v p="$pane" '$1!=p' "$LAST_RESPONSE_FILE" > "$tmp"
+  printf '%s\t%s\t%s\t%s\t%s\n' "$pane" "$now" "$state" "$action" "$clean_prompt" >> "$tmp"
+  mv "$tmp" "$LAST_RESPONSE_FILE"
+}
+
+prior_answer_for() {
+  local pane="$1"
+  awk -F'\t' -v p="$pane" -v now="$(date +%s)" '
+    $1==p {
+      age = now - $2
+      printf "(%ds ago, state=%s, action=%s)\n  prompt: %s\n", age, $3, $4, $5
+      exit
+    }
+  ' "$LAST_RESPONSE_FILE" 2>/dev/null
+}
+
+# ── metrics: per-decision TSV with next-tick outcome backfill ───────────────
+# Columns: ts | pane | panel | state | model | action | rc | outcome
+# outcome starts as "?" and is rewritten by the NEXT tick for this pane:
+#   resolved   → pane is now busy/quiet (the paste worked)
+#   unresolved → pane is still asking/blocked (paste didn't unstick it)
+# Combined with the strike guard this is the foundation for evaluating
+# whether tiered models / recommended-path heuristic actually help.
+METRICS_FILE="$STATE_DIR/metrics.tsv"
+[ -f "$METRICS_FILE" ] || : > "$METRICS_FILE"
+
+record_metric() {
+  local pane="$1" panel="$2" state="$3" model="$4" action="$5" rc="$6"
+  local now; now="$(date +%s)"
+  printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\t?\n' \
+    "$now" "$pane" "$panel" "$state" "$model" "$action" "$rc" >> "$METRICS_FILE"
+}
+
+# For the most-recent row of `pane` whose outcome column is "?", rewrite
+# it to `verdict`. Called on a pane's NEXT tick: if pane is now quiet/busy
+# the previous decision is "resolved"; if still asking/blocked it's
+# "unresolved". No-op when no prior row or no "?" outcome — keeps the
+# script cheap on cold panes.
+backfill_outcome() {
+  local pane="$1" verdict="$2"
+  awk -F'\t' -v OFS='\t' -v p="$pane" -v v="$verdict" '
+    BEGIN { found=0 }
+    {
+      if (!found && $2==p && $8=="?") {
+        $8 = v
+        found = 1
+      }
+      print
+    }
+  ' "$METRICS_FILE" > "$METRICS_FILE.tmp" && mv "$METRICS_FILE.tmp" "$METRICS_FILE"
+}
+
+# Translate a state name into a metrics outcome label.
+state_to_outcome() {
+  case "$1" in
+    busy|quiet)     echo resolved ;;
+    asking|blocked) echo unresolved ;;
+    *)              echo unknown ;;
+  esac
+}
+
+# Count strikes for a pane within STRIKE_WINDOW seconds.
+strike_count() {
+  local pane="$1" now="$2"
+  awk -F'\t' -v p="$pane" -v cut="$((now - STRIKE_WINDOW))" \
+    '$2==p && $1+0 >= cut {n++} END{print n+0}' "$PASTE_LOG"
+}
+
+record_paste() {
+  local pane="$1" state="$2" now; now="$(date +%s)"
+  printf '%s\t%s\t%s\n' "$now" "$pane" "$state" >> "$PASTE_LOG"
+}
+
+record_escalation() {
+  local pane="$1" panel="$2" state="$3" strikes="$4" now; now="$(date +%s)"
+  printf '%s\t%s\t%s\t%s\tstrikes=%d in %ds\n' \
+    "$now" "$pane" "$panel" "$state" "$strikes" "$STRIKE_WINDOW" >> "$ESCALATE_LOG"
+}
+
+within_cooldown() {
+  local pane="$1" now="$2"
+  local last; last="$(awk -F'\t' -v p="$pane" '$1==p {print $2; exit}' "$STATE_FILE")"
+  [ -z "$last" ] && return 1
+  [ "$((now - last))" -lt "$COOLDOWN" ]
+}
+
+already_answered() {
+  local pane="$1" hash="$2"
+  awk -F'\t' -v p="$pane" -v h="$hash" '$1==p && $3==h {f=1} END{exit !f}' "$STATE_FILE"
+}
+
+record() {
+  local pane="$1" hash="$2" now; now="$(date +%s)"
+  local tmp; tmp="$(mktemp)"
+  awk -F'\t' -v p="$pane" '$1!=p' "$STATE_FILE" > "$tmp"
+  printf '%s\t%s\t%s\n' "$pane" "$now" "$hash" >> "$tmp"
+  mv "$tmp" "$STATE_FILE"
+}
+
+# ── plan / queue context ────────────────────────────────────────────────────
+active_plan() {
+  cat "$REPO_ROOT/.codex-fleet/active-plan" 2>/dev/null || true
+}
+
+plan_summary() {
+  # One-line per plan: slug | done | claimed | available
+  local slug="$1"
+  local pj="$REPO_ROOT/openspec/plans/$slug/plan.json"
+  [ -f "$pj" ] || { echo "(plan.json missing: $slug)"; return; }
+  jq -r --arg s "$slug" '
+    [.tasks[] | (.status // "open")] as $st |
+    "\($s)\tdone=\($st | map(select(.=="complete" or .=="done" or .=="completed" or .=="merged")) | length)\tclaimed=\($st | map(select(.=="claimed")) | length)\tavailable=\($st | map(select(.=="available")) | length)"
+  ' "$pj"
+}
+
+all_plans_summary() {
+  for slug in "$REPO_ROOT"/openspec/plans/*/; do
+    local s; s="$(basename "$slug")"
+    plan_summary "$s"
+  done
+}
+
+# Title + problem + acceptance criteria of the active plan. The
+# build_prompt's "pick recommended unless it breaks the plan" rule has no
+# teeth without this — claude can't audit a choice against criteria it
+# never saw. Capped so a verbose plan can't blow out the context budget.
+PLAN_CONTENT_BUDGET="${PLAN_CONTENT_BUDGET:-2400}"   # chars
+active_plan_content() {
+  local slug="$1"
+  [ -z "$slug" ] && return 0
+  local pj="$REPO_ROOT/openspec/plans/$slug/plan.json"
+  [ -f "$pj" ] || return 0
+  jq -r '
+    "TITLE: " + (.title // "(no title)"),
+    "",
+    "PROBLEM:",
+    (.problem // "(no problem statement)"),
+    "",
+    "ACCEPTANCE CRITERIA:",
+    ((.acceptance_criteria // []) | to_entries | map("  " + ((.key+1)|tostring) + ". " + .value) | join("\n"))
+  ' "$pj" 2>/dev/null | head -c "$PLAN_CONTENT_BUDGET"
+}
+
+# ── build the supervisor prompt (user message only) ────────────────────────
+# The role / decision rules / output schema live in the cached system
+# prompt ($SYSTEM_PROMPT_FILE). Only the per-call variable bits go here.
+build_prompt() {
+  local pane="$1" panel="$2" tail="$3" plan="$4" state="$5"
+  cat <<HEAD
+Pane: $panel (tmux id $pane)
+State: $state
+
+# Active plan pinned in .codex-fleet/active-plan
+$plan
+
+# Active plan content (title + problem + acceptance criteria) — audit the
+# recommended option against these before picking.
+$(active_plan_content "$plan" | sed 's/^/  /')
+
+# Plan registry — counts of done / claimed / available sub-tasks
+$(all_plans_summary | sed 's/^/  /')
+
+# Previously you told this pane (most recent first, may be empty)
+$(prior_answer_for "$pane" | sed 's/^/  /')
+
+# Recent worker output (last 80 lines, ANSI stripped)
+\`\`\`
+$(printf '%s\n' "$tail" | sed -E 's/\x1b\[[0-9;]*[a-zA-Z]//g')
+\`\`\`
+HEAD
+}
+
+# ── apply claude's response ─────────────────────────────────────────────────
+# Tries JSON parse first (schema-validated path). If that fails — e.g. the
+# fallback model ignored the schema, or the CLI returned a JSON envelope
+# with the schema'd object nested inside — falls back to the legacy line
+# parser so we don't lose the response. Either way ends with the same
+# (action, plan_line, tool, prompt) locals.
+apply_response() {
+  local pane="$1" panel="$2" response="$3" state="${4:-?}"
+  local action plan_line prompt tool source="json"
+
+  if printf '%s' "$response" | jq -e . >/dev/null 2>&1; then
+    action="$(printf '%s' "$response"    | jq -r '.action  // ""')"
+    plan_line="$(printf '%s' "$response" | jq -r '.plan    // ""')"
+    tool="$(printf '%s' "$response"      | jq -r '.tool    // ""')"
+    prompt="$(printf '%s' "$response"    | jq -r '.prompt  // ""')"
+  else
+    source="text-fallback"
+    action="$(printf '%s\n' "$response"    | grep -m1 -iE '^(ACTION|"action")' | sed -E 's/^[^:]*: *//; s/^"//; s/",?$//' | awk '{print $1}')"
+    plan_line="$(printf '%s\n' "$response" | grep -m1 -iE '^(PLAN|"plan")'     | sed -E 's/^[^:]*: *//; s/^"//; s/",?$//')"
+    tool="$(printf '%s\n' "$response"      | grep -m1 -iE '^(TOOL|"tool")'     | sed -E 's/^[^:]*: *//; s/^"//; s/",?$//')"
+    prompt="$(printf '%s\n' "$response"    | awk 'BEGIN{IGNORECASE=1} /^(PROMPT|"prompt")/ {flag=1; sub(/^[^:]*: */,""); sub(/^"/,""); sub(/",?$/,""); print; next} flag')"
+  fi
+  log "  $panel: action=${action:-?} plan=${plan_line:-(none)} tool=${tool:-(none)} via=$source"
+
+  if [ "$action" = "C" ] || [ -z "$prompt" ] || [ "$prompt" = "(none)" ]; then
+    log "  $panel: no action — letting the worker keep working"
+    return 2   # caller uses rc==2 to skip strike accounting
+  fi
+
+  if [ "$DRY_RUN" = "1" ]; then
+    log "  $panel: DRY-RUN — would paste $(printf '%s' "$prompt" | wc -c) chars"
+    printf '    %s\n' "$(printf '%s' "$prompt" | head -c 200)" | sed 's/^/    /'
+    return 0
+  fi
+
+  local buf="claude-sup-$$-${pane#%}"
+  printf '%s' "$prompt" | tmux load-buffer -b "$buf" -
+  tmux send-keys -t "$pane" Escape 2>/dev/null || true
+  tmux paste-buffer -b "$buf" -t "$pane" -p
+  tmux send-keys -t "$pane" Enter
+  tmux delete-buffer -b "$buf" 2>/dev/null || true
+  log "  $panel: pasted response"
+  record_last_response "$pane" "$state" "${action:-?}" "$prompt"
+  return 0   # rc==0 → tick() will record a strike
+}
+
+# ── one tick ────────────────────────────────────────────────────────────────
+tick() {
+  local now; now="$(date +%s)"
+  local plan; plan="$(active_plan)"
+  [ -z "$plan" ] && plan="(no plan pinned)"
+
+  local panes
+  if [ -n "$ONLY_PANE" ]; then
+    panes="$ONLY_PANE|$(tmux show-option -p -t "$ONLY_PANE" '@panel' 2>/dev/null | awk '{print $2}' | tr -d '"')"
+  else
+    panes="$(tmux list-panes -t "$SESSION:$WINDOW" -F '#{pane_id}|#{@panel}' 2>/dev/null)"
+  fi
+
+  printf '%s\n' "$panes" | while IFS='|' read -r pane panel; do
+    [ -z "$panel" ] && continue
+    [ "$panel" = "[codex-fleet-tab-strip]" ] && continue
+
+    local tail; tail="$(tmux capture-pane -p -t "$pane" -S -80 2>/dev/null || true)"
+    local state; state="$(classify_tail "$tail")"
+    # Backfill the previous decision's outcome BEFORE we early-return —
+    # even "this pane is now busy, skip" is a valid datapoint (it tells
+    # us the previous paste worked).
+    backfill_outcome "$pane" "$(state_to_outcome "$state")"
+    case "$state" in
+      busy|quiet) continue ;;
+      asking|blocked) : ;;
+    esac
+    within_cooldown "$pane" "$now" && { log "$panel: cooldown — skip"; continue; }
+
+    local h; h="$(tail_hash "$tail")"
+    if already_answered "$pane" "$h"; then
+      log "$panel: already answered this exact state (hash=$h) — skip"
+      continue
+    fi
+
+    # 3-strike loop guard: if we've pasted to this pane >= STRIKE_LIMIT
+    # times inside the trailing window, the worker is in a loop the
+    # supervisor isn't getting out of — escalate to the operator instead
+    # of paying for another claude turn.
+    local strikes; strikes="$(strike_count "$pane" "$now")"
+    if [ "$strikes" -ge "$STRIKE_LIMIT" ]; then
+      log "$panel: $state — strike guard tripped ($strikes/$STRIKE_LIMIT in ${STRIKE_WINDOW}s) — escalating"
+      record_escalation "$pane" "$panel" "$state" "$strikes"
+      record "$pane" "$h"   # mark hash so we don't keep tripping every tick
+      continue
+    fi
+
+    # Pick model tier by state. asking = sonnet/medium (fast, cheap),
+    # blocked = opus-4-7/high (real reasoning). Both share the JSON
+    # schema so apply_response sees the same shape either way.
+    local -a args; mapfile -t args < <(claude_args_for "$state")
+    log "$panel: $state — asking ${args[*]} (strikes=$strikes/$STRIKE_LIMIT)"
+    local prompt_txt; prompt_txt="$(build_prompt "$pane" "$panel" "$tail" "$plan" "$state")"
+    local response
+    # Resolve the tier model for the metrics row.
+    local tier_model
+    case "$state" in
+      asking) tier_model="${CLAUDE_MODEL_ASKING:-sonnet}" ;;
+      *)      tier_model="${CLAUDE_MODEL_BLOCKED:-claude-opus-4-7}" ;;
+    esac
+    if response="$("$CLAUDE_BIN" -p "${args[@]}" <<<"$prompt_txt" 2>>"$STATE_DIR/claude.err")"; then
+      local rc=0
+      apply_response "$pane" "$panel" "$response" "$state" || rc=$?
+      record "$pane" "$h"
+      [ "$rc" = "0" ] && record_paste "$pane" "$state"
+      # Best-effort action extraction for the metrics row. apply_response
+      # already parsed it but we don't return it, so re-extract cheaply.
+      local m_action
+      if printf '%s' "$response" | jq -e . >/dev/null 2>&1; then
+        m_action="$(printf '%s' "$response" | jq -r '.action // "?"')"
+      else
+        m_action="$(printf '%s\n' "$response" | grep -m1 -iE '^(ACTION|"action")' | sed -E 's/^[^:]*: *//; s/^"//; s/",?$//' | awk '{print $1}')"
+      fi
+      record_metric "$pane" "$panel" "$state" "$tier_model" "${m_action:-?}" "$rc"
+    else
+      log "  $panel: claude -p failed — see $STATE_DIR/claude.err"
+      record_metric "$pane" "$panel" "$state" "$tier_model" "ERR" "1"
+    fi
+  done
+}
+
+# ── main ────────────────────────────────────────────────────────────────────
+if (( ONCE == 1 )); then
+  tick
+  exit 0
+fi
+
+log "loop start (interval=${INTERVAL}s cooldown=${COOLDOWN}s min-blocked=${MIN_BLOCKED}s)"
+while :; do
+  tick || log "tick error — continuing"
+  sleep "$INTERVAL"
+done
diff --git a/scripts/codex-fleet/lib/claude-supervisor-classifier.sh b/scripts/codex-fleet/lib/claude-supervisor-classifier.sh
new file mode 100644
index 0000000..1b38ac1
--- /dev/null
+++ b/scripts/codex-fleet/lib/claude-supervisor-classifier.sh
@@ -0,0 +1,228 @@
+#!/usr/bin/env bash
+# claude-supervisor-classifier.sh — pure pane-tail classifier extracted from
+# claude-supervisor.sh so the daemon and the replay harness share one
+# implementation. This file is sourced; it MUST NOT call tmux, claude,
+# date(1) for state, or write any files at source time.
+#
+# Public functions:
+#   classify_tail "<tail>"   → echoes one of: busy | asking | blocked | quiet
+#   is_busy / is_asking / is_blocked / last_line_is_prompt
+#   tail_hash "<tail>"       → 12-char stable hash with timestamps stripped
+#
+# Public arrays (read-only after source):
+#   BUSY_PATTERNS, ASK_PATTERNS, BLOCKED_PATTERNS
+
+# How many trailing non-empty lines count as "recent" for the busy/ask gates.
+# 80 is the full capture window; 8 is roughly the on-screen visible region of
+# a codex pane after ANSI is stripped. Stale busy/ask signals further back in
+# scrollback do not represent the worker's current state, so we ignore them.
+: "${CLAUDE_SUPERVISOR_RECENT_LINES:=8}"
+
+# Patterns that mean the worker is genuinely blocked and needs help — not
+# just "polling for the next ready task". Matched with grep -iE on the full
+# tail (blockers are sticky; a stale blocker the worker hasn't unstuck is
+# still a real blocker).
+BLOCKED_PATTERNS=(
+  "PLAN_SUBTASK_NOT_FOUND"
+  "PLAN_SUBTASK_NOT_AVAILABLE"
+  "stale.*blocker"
+  "stale-claim"
+  "did not rescue"
+  "told me not to rescue"
+  "No branch, worktree, edits, or PR were created"
+  "BLOCKED preflight"
+  "BLOCKED writable-root"
+  "I waited and rechecked"
+  "less than 5% of your 5h limit"
+  "^.*Blocked\."
+  # Added 2026-05-15 — real failure modes seen in the wild that the
+  # supervisor was deaf to:
+  "CONFLICT \(content"             # git merge conflict
+  "merge conflict"                  # plain-text variant
+  "Your local changes .* would be overwritten"
+  "error: uncommitted changes"
+  "^fatal: "                        # git fatal
+  "Permission denied \(publickey\)" # ssh auth bust
+  "gh: command not found"
+  "gh auth: "                       # gh auth error variants
+  "Bad credentials"                 # GitHub token rejected
+  "MCP server .*(not found|missing|unavailable)"
+  "429 Too Many Requests"           # cap-swap should catch this, but in
+                                    # case it doesn't, surface to supervisor
+  "remote: Permission to .* denied"
+  "BLOCKED:"                        # canonical Guardex blocker prefix
+)
+
+# Patterns that mean codex is talking BACK — a question, a menu, an approval
+# prompt. These are matched on a *narrow* trailing window combined with the
+# last_line_is_prompt cursor gate so that mid-task narration ("Should I
+# also lint?") does not trip the classifier.
+ASK_PATTERNS=(
+  "\(recommended\)"
+  "\(default\)"
+  "\[Y/n\]"
+  "\[y/N\]"
+  "\(y/n\)"
+  "Continue\?"
+  "Approve\?"
+  "Proceed\?"
+  "Confirm\?"
+  "Apply\?"
+  "Would you like"
+  "Do you want"
+  "Should I"
+  "Which option"
+  "Which approach"
+  "Which one"
+  "Choose one"
+  "Choose:"
+  "Select one"
+  "Select:"
+  "Pick one"
+  "press [0-9] to"
+  "type [0-9] to"
+  "Need clarification"
+  "Need more (info|context|detail)"
+  "Please (clarify|confirm|choose|specify)"
+  "^[[:space:]]*[1-9][.)] .*\?$"
+  "^[[:space:]]*[1-9][.)] .*\(recommended\)"
+)
+
+# Patterns that mean the worker is actively executing — DO NOT interrupt.
+# Matched only on the recent window so a 40-line-old `Working (12s)` doesn't
+# mask a brand-new `[Y/n]` cursor on the bottom line.
+BUSY_PATTERNS=(
+  "esc to interrupt"
+  "Working ("
+  # codex's post-completion footer reads "Worked for 5m 12s …". That line
+  # by itself means the worker is now IDLE (quiet), not busy — leave it
+  # out of BUSY_PATTERNS. The plan-watcher fast path picks it up.
+)
+
+# Strip ANSI then return the last N non-empty lines, newline-separated.
+# Used by is_busy and is_asking to scope detection to the recent on-screen
+# region. N defaults to CLAUDE_SUPERVISOR_RECENT_LINES.
+tail_recent_lines() {
+  local tail="$1" n="${2:-$CLAUDE_SUPERVISOR_RECENT_LINES}"
+  printf '%s\n' "$tail" \
+    | sed -E 's/\x1b\[[0-9;?]*[a-zA-Z]//g' \
+    | awk 'NF' \
+    | tail -n "$n"
+}
+
+# Busy detection is anchored to the LAST non-empty line. codex's CLI rewrites
+# the "Working (12s • ↑ 4.2k tokens • esc to interrupt)" footer in place; if
+# the worker is genuinely busy that line is the bottom of the capture. Once
+# the turn completes, codex replaces it with a "Worked for …" footer and the
+# pane is no longer busy. Looking deeper than the last line lets a stale
+# `Working (` in scrollback mask a brand-new interactive menu below it
+# (FM-B in the audit).
+is_busy() {
+  local tail="$1"
+  local last; last="$(printf '%s\n' "$tail" \
+    | sed -E 's/\x1b\[[0-9;?]*[a-zA-Z]//g' \
+    | awk 'NF{ln=$0} END{print ln}')"
+  [ -z "$last" ] && return 1
+  local pat
+  for pat in "${BUSY_PATTERNS[@]}"; do
+    printf '%s\n' "$last" | grep -qF -- "$pat" && return 0
+  done
+  return 1
+}
+
+is_blocked() {
+  local tail="$1"
+  local pat
+  for pat in "${BLOCKED_PATTERNS[@]}"; do
+    if printf '%s\n' "$tail" | grep -qiE -- "$pat"; then return 0; fi
+  done
+  return 1
+}
+
+is_asking() {
+  local tail="$1"
+  # Two-gate precision filter:
+  #  1) the bottom of the pane must LOOK like a waiting cursor
+  #     (last_line_is_prompt)
+  #  2) one of the ASK_PATTERNS must appear in the same recent window
+  # Either gate alone produces false positives — scrollback `Continue?`
+  # 60 lines back + a normal status line ending in `:` previously slipped
+  # through. Both gates together cut FPs to ~0 on our fixture corpus.
+  last_line_is_prompt "$tail" || return 1
+  local recent; recent="$(tail_recent_lines "$tail")"
+  local pat
+  for pat in "${ASK_PATTERNS[@]}"; do
+    if printf '%s\n' "$recent" | grep -qiE -- "$pat"; then return 0; fi
+  done
+  return 1
+}
+
+# True iff the bottom of the captured tail looks like codex is sitting at a
+# waiting prompt. We accept ONLY high-precision cursor shapes — a bare
+# trailing `:` no longer qualifies (workers narrate "Reading file:" all the
+# time), and a bare trailing `?` only qualifies when the same line carries a
+# known question lead-word.
+last_line_is_prompt() {
+  local tail="$1"
+  local last; last="$(printf '%s\n' "$tail" \
+    | sed -E 's/\x1b\[[0-9;?]*[a-zA-Z]//g' \
+    | awk 'NF{ln=$0} END{print ln}')"
+  [ -z "$last" ] && return 1
+
+  # High-precision cursor shapes — these only appear at the bottom when the
+  # CLI is genuinely waiting. Bare `:$` was previously here and was the
+  # #1 false-positive source; it is removed deliberately.
+  local cursor_patterns=(
+    '\[Y/n\][[:space:]]*$'
+    '\[y/N\][[:space:]]*$'
+    '\(y/n\)[[:space:]]*$'
+    '^[[:space:]]*[>❯▶➜▷»][[:space:]]*$'
+    '[>❯▶➜▷»][[:space:]]*$'
+    '\$[[:space:]]*$'
+    '#[[:space:]]*$'
+    'press [0-9]'
+    'type [0-9]'
+    'press [a-z] to'
+    '^[[:space:]]*[0-9]+[).][[:space:]]+.*\(recommended\)'
+    '^[[:space:]]*[0-9]+[).][[:space:]]+.*\(default\)'
+  )
+  local pat
+  for pat in "${cursor_patterns[@]}"; do
+    if printf '%s\n' "$last" | grep -qiE -- "$pat"; then return 0; fi
+  done
+
+  # Bare trailing `?` is admitted only when the same line carries a known
+  # question lead-word. Catches "Continue?", "Approve?", "Should I … ?"
+  # but rejects "do you want to keep going? yes." in mid-paragraph
+  # narration that happens to wrap onto the last line of the capture.
+  case "$last" in
+    *\?|*\?\ )
+      printf '%s\n' "$last" | grep -qiE -- '(Continue|Approve|Proceed|Confirm|Apply|Should I|Do you want|Would you like|Which (option|approach|one)|Choose|Select|Pick|Need (clarification|more)|Please (clarify|confirm|choose|specify))'
+      return $?
+      ;;
+  esac
+  return 1
+}
+
+# Classify the recent tail into busy | asking | blocked | quiet. Ordering
+# matters: `asking` outranks `blocked` because a pane that mentioned a stale
+# blocker but is now showing an interactive menu wants an answer to the
+# menu, not a rescue lecture. `busy` is only checked against the recent
+# window now, so a fresh ask under a stale "Working (" no longer slips.
+classify_tail() {
+  local tail="$1"
+  if is_busy    "$tail"; then echo busy;    return; fi
+  if is_asking  "$tail"; then echo asking;  return; fi
+  if is_blocked "$tail"; then echo blocked; return; fi
+  echo quiet
+}
+
+# Hash of the relevant tail so we don't re-ask claude the same question.
+# Strip ANSI + timestamps so "Worked for 9m 36s → 10m 12s" or wall-clock
+# ticking doesn't bust the cache when nothing real changed.
+tail_hash() {
+  local tail="$1"
+  printf '%s\n' "$tail" \
+    | sed -E 's/\x1b\[[0-9;]*[a-zA-Z]//g; s/Worked for [0-9smh ]+//g; s/[0-9]{2}:[0-9]{2}:[0-9]{2}//g' \
+    | sha1sum | cut -c1-12
+}
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__busy_in_scrollback_fresh_cursor.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__busy_in_scrollback_fresh_cursor.txt
new file mode 100644
index 0000000..dcea099
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__busy_in_scrollback_fresh_cursor.txt
@@ -0,0 +1,11 @@
+> claim sub-task 1 of plan codex-fleet-glass-menu-drop-tabstrip-2026-05-15
+  reading plan.json
+Working (8s • ↑ 1.2k tokens • esc to interrupt)
+Worked for 6m 32s • ↑ 12.4k tokens
+> file rust/fleet-ui/src/glass/mod.rs already exists, contents diverge from plan
+
+Detected merge of generated migration vs hand-edited one — which approach do you prefer?
+  1) accept generated (recommended)
+  2) keep hand-edited
+  3) abort the sub-task
+❯
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__press_digit_to_continue.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__press_digit_to_continue.txt
new file mode 100644
index 0000000..2680288
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__press_digit_to_continue.txt
@@ -0,0 +1,6 @@
+  Detected multiple matching accounts:
+    1) acc-claude-pro-1@example.com
+    2) acc-claude-pro-2@example.com
+    3) acc-claude-team-1@example.com
+
+press 1 to pick acc-claude-pro-1, 2 for pro-2, 3 for team-1
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__recommended_numbered_menu.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__recommended_numbered_menu.txt
new file mode 100644
index 0000000..efcd254
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__recommended_numbered_menu.txt
@@ -0,0 +1,7 @@
+  Detected three candidate approaches for this refactor:
+  1) inline the helper into the only call site
+  2) extract a new module under rust/fleet-ui/src/glass/ (recommended)
+  3) leave the duplication and tag with TODO
+
+  Which option would you like me to take?
+2) extract a new module under rust/fleet-ui/src/glass/ (recommended)
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__stale_blocker_in_scrollback_menu_now.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__stale_blocker_in_scrollback_menu_now.txt
new file mode 100644
index 0000000..c1b7955
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__stale_blocker_in_scrollback_menu_now.txt
@@ -0,0 +1,11 @@
+  earlier in this lane:
+  > BLOCKED: stale-claim blocker on sub-task 2, told me not to rescue
+  > queue is empty, what now?
+  operator: claim was rescued, retry plan-watcher pulled a fresh sub-task
+
+  resumed: sub-task 5 of codex-fleet-ios-pages-design-pass-2026-05-15
+  reading plan content
+  Detected migration drift, two options:
+  1) regenerate the migration (recommended)
+  2) hand-edit the current migration
+1) regenerate the migration (recommended)
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__yn_lowercase_default.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__yn_lowercase_default.txt
new file mode 100644
index 0000000..9d87aab
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__yn_lowercase_default.txt
@@ -0,0 +1,4 @@
+  Running `gx branch finish --branch agent/claude/x --via-pr --wait-for-merge`.
+  This will push the branch and open a PR.
+
+Continue? [Y/n]
\ No newline at end of file
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__yn_uppercase_destructive.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__yn_uppercase_destructive.txt
new file mode 100644
index 0000000..09c2378
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/asking__yn_uppercase_destructive.txt
@@ -0,0 +1,4 @@
+  About to run: rm -rf /tmp/codex-fleet/sandboxes/agent-claude-old-2026-05-13
+  This is irreversible.
+
+Proceed? [y/N]
\ No newline at end of file
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__bad_credentials.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__bad_credentials.txt
new file mode 100644
index 0000000..4d3f6a4
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__bad_credentials.txt
@@ -0,0 +1,4 @@
+> gh pr create --title "feat(...)"
+  error: Bad credentials
+  authentication failed for github.com — token may be stale
+  refusing to open PR
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__fatal_git.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__fatal_git.txt
new file mode 100644
index 0000000..e53bbf0
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__fatal_git.txt
@@ -0,0 +1,3 @@
+> git push -u origin agent/claude/glass-menu-drop-2026-05-15
+fatal: unable to access 'https://github.com/recodeee/codex-fleetui.git/': Could not resolve host: github.com
+  pushing aborted
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__five_pct_limit.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__five_pct_limit.txt
new file mode 100644
index 0000000..a2bc65b
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__five_pct_limit.txt
@@ -0,0 +1,5 @@
+> mid-task: editing rust/fleet-waves/src/main.rs
+  edits queued
+  attempting to send to model
+  warning: less than 5% of your 5h limit remaining
+  pausing — will not start another long turn
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__mcp_server_missing.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__mcp_server_missing.txt
new file mode 100644
index 0000000..a175c04
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__mcp_server_missing.txt
@@ -0,0 +1,4 @@
+> mcp__colony__task_ready_for_agent
+Error: MCP server colony not found in the registered servers
+  available: claude-mem, gbrain
+  cannot continue — Colony required by plan
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__merge_conflict.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__merge_conflict.txt
new file mode 100644
index 0000000..95d8ad4
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__merge_conflict.txt
@@ -0,0 +1,6 @@
+> gx branch finish --branch agent/claude/x --via-pr --wait-for-merge --cleanup
+  merging into main...
+  Auto-merging rust/fleet-ui/src/rail.rs
+  CONFLICT (content): Merge conflict in rust/fleet-ui/src/rail.rs
+  Automatic merge failed; fix conflicts and commit the result.
+Worked for 7m 02s
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__permission_denied.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__permission_denied.txt
new file mode 100644
index 0000000..20f29a0
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__permission_denied.txt
@@ -0,0 +1,4 @@
+> git push -u origin agent/claude/cap-swap-2026-05-15
+ERROR: Permission denied (publickey).
+fatal: Could not read from remote repository.
+Please make sure you have the correct access rights.
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__plan_subtask_not_found.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__plan_subtask_not_found.txt
new file mode 100644
index 0000000..f63bf60
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__plan_subtask_not_found.txt
@@ -0,0 +1,5 @@
+> OVERRIDE: claim sub-task 0 of plan codex-fleet-glass-menu-drop-tabstrip-2026-05-15
+  reading openspec/plans/codex-fleet-glass-menu-drop-tabstrip-2026-05-15/plan.json
+  PLAN_SUBTASK_NOT_FOUND: sub-task 0 is already complete, no available work
+  queue is empty
+Blocked. PLAN_SUBTASK_NOT_FOUND, queue empty.
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__stale_claim_told_not_rescue.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__stale_claim_told_not_rescue.txt
new file mode 100644
index 0000000..fcccacb
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__stale_claim_told_not_rescue.txt
@@ -0,0 +1,5 @@
+> task_ready_for_agent: nothing ready
+  found stale-claim blocker on sub-task 2 of codex-fleet-ios-pages-design-pass-2026-05-15
+  operator-policy: told me not to rescue stale claims unsupervised
+  posted BLOCKED handoff, exiting
+Worked for 14m 02s • ↑ 18.3k tokens
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__uncommitted_changes.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__uncommitted_changes.txt
new file mode 100644
index 0000000..e9fe113
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/blocked__uncommitted_changes.txt
@@ -0,0 +1,4 @@
+> git pull --ff-only
+  error: uncommitted changes in 4 files
+  hint: stash, finish in worktree, or commit before pulling
+  refusing to proceed
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/busy__codex_working_active.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/busy__codex_working_active.txt
new file mode 100644
index 0000000..4b57595
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/busy__codex_working_active.txt
@@ -0,0 +1,6 @@
+> claim sub-task 0 of plan codex-fleet-glass-menu-drop-tabstrip-2026-05-15
+  reading openspec/plans/codex-fleet-glass-menu-drop-tabstrip-2026-05-15/plan.json
+  picked sub-task 0: glass-menu-base-layer
+  starting agent worktree
+  edited 3 files
+Working (24s • ↑ 4.2k tokens • esc to interrupt)
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/busy__esc_to_interrupt_last_line.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/busy__esc_to_interrupt_last_line.txt
new file mode 100644
index 0000000..c7b3699
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/busy__esc_to_interrupt_last_line.txt
@@ -0,0 +1,5 @@
+> running cargo test -p fleet-tab-strip
+  Compiling fleet-tab-strip v0.1.0
+  Compiling fleet-ui v0.1.0
+test result: ok. 18 passed; 0 failed
+Working (1m 12s • ↑ 7.8k tokens • esc to interrupt)
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__colon_ending_stale_continue.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__colon_ending_stale_continue.txt
new file mode 100644
index 0000000..7644195
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__colon_ending_stale_continue.txt
@@ -0,0 +1,10 @@
+  long-running task started
+  > Continue? yes — answered earlier
+  > running tests
+  Compiling fleet-data
+  Compiling fleet-state
+  Compiling fleet-launcher
+  cargo test -p fleet-launcher: 12 passed
+  ran scripts/codex-fleet/lib/sandboxed-cargo-test.sh: ok
+  next:
+Reading file: rust/fleet-launcher/src/heartbeat.rs:
\ No newline at end of file
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__do_you_want_in_narration.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__do_you_want_in_narration.txt
new file mode 100644
index 0000000..8efcc66
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__do_you_want_in_narration.txt
@@ -0,0 +1,6 @@
+  reviewed the plan: the operator wrote "do you want me to keep tabs visible? no, drop them"
+  honoring that — dropping the tab strip in the glass layer
+  applied diff to rust/fleet-tab-strip/src/main.rs
+  cargo test -p fleet-tab-strip: 18 passed
+  gx branch finish queued
+Worked for 6m 41s • ↑ 9.0k tokens
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__empty_pane.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__empty_pane.txt
new file mode 100644
index 0000000..139597f
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__empty_pane.txt
@@ -0,0 +1,2 @@
+
+
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__narrative_should_i_no_cursor.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__narrative_should_i_no_cursor.txt
new file mode 100644
index 0000000..837f075
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__narrative_should_i_no_cursor.txt
@@ -0,0 +1,8 @@
+> I'll start by running the migration. Should I lint after? Yes — running pnpm lint now.
+  Compiling fleet-ui
+  pnpm lint: 0 errors
+  next: cargo test -p fleet-ui
+  cargo test -p fleet-ui: 24 passed
+  next: gx branch finish
+  ran finish via-pr, PR opened, merged
+Worked for 4m 18s • ↑ 6.1k tokens
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__numbered_list_narrative_summary.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__numbered_list_narrative_summary.txt
new file mode 100644
index 0000000..a5909c3
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__numbered_list_narrative_summary.txt
@@ -0,0 +1,7 @@
+  Summary of what I did this sub-task:
+  1) extracted the glass layer to its own module
+  2) added the new palette entries
+  3) wired up the tab-strip drop
+  4) ran cargo test -p fleet-tab-strip
+  PR opened, merged.
+Worked for 9m 02s • ↑ 13.1k tokens
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__prompt_ready_for_input.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__prompt_ready_for_input.txt
new file mode 100644
index 0000000..6639f43
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__prompt_ready_for_input.txt
@@ -0,0 +1,3 @@
+  plan-watcher: idle since 5m 02s, waiting for next available sub-task
+  no plan currently pinned in .codex-fleet/active-plan
+  >
diff --git a/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__worked_for_completion_footer.txt b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__worked_for_completion_footer.txt
new file mode 100644
index 0000000..0352b8e
--- /dev/null
+++ b/scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/quiet__worked_for_completion_footer.txt
@@ -0,0 +1,5 @@
+> claim sub-task 0 of plan codex-fleet-glass-menu-drop-tabstrip-2026-05-15
+  edits made to rust/fleet-tab-strip/src/main.rs
+  cargo build -p fleet-tab-strip: ok
+  gx branch finish: PR #128 opened, merged
+Worked for 5m 12s • ↑ 8.4k tokens
diff --git a/scripts/codex-fleet/test/test-claude-supervisor-classifier.sh b/scripts/codex-fleet/test/test-claude-supervisor-classifier.sh
new file mode 100755
index 0000000..44ba1ef
--- /dev/null
+++ b/scripts/codex-fleet/test/test-claude-supervisor-classifier.sh
@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+# test-claude-supervisor-classifier.sh — fixture-driven replay harness for
+# the claude-supervisor classifier. Sources the pure lib at
+# scripts/codex-fleet/lib/claude-supervisor-classifier.sh and runs
+# classify_tail against every fixture under
+# scripts/codex-fleet/test/fixtures/claude-supervisor-classifier/.
+#
+# Fixture naming:
+#   <expected-label>__<short-name>.txt
+# where <expected-label> is one of: busy | asking | blocked | quiet.
+#
+# Exit codes:
+#   0 — every fixture matched its expected label
+#   1 — at least one fixture mismatched
+#   2 — usage error / missing lib
+
+set -u
+set -o pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+FLEET_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+LIB="$FLEET_DIR/lib/claude-supervisor-classifier.sh"
+FIXTURE_DIR="${1:-$SCRIPT_DIR/fixtures/claude-supervisor-classifier}"
+
+if [ ! -f "$LIB" ]; then
+  printf 'fatal: classifier lib not found at %s\n' "$LIB" >&2
+  exit 2
+fi
+if [ ! -d "$FIXTURE_DIR" ]; then
+  printf 'fatal: fixture dir not found at %s\n' "$FIXTURE_DIR" >&2
+  exit 2
+fi
+
+# shellcheck source=../lib/claude-supervisor-classifier.sh
+. "$LIB"
+
+pass=0
+fail=0
+fails=()
+
+shopt -s nullglob
+for fixture in "$FIXTURE_DIR"/*.txt; do
+  base="$(basename "$fixture")"
+  expected="${base%%__*}"
+  case "$expected" in
+    busy|asking|blocked|quiet) ;;
+    *)
+      printf 'skip: %s (bad prefix, expected busy|asking|blocked|quiet)\n' "$base"
+      continue
+      ;;
+  esac
+
+  tail_content="$(cat "$fixture")"
+  actual="$(classify_tail "$tail_content")"
+
+  if [ "$actual" = "$expected" ]; then
+    pass=$((pass+1))
+    printf '  PASS  %-12s  %s\n' "$expected" "$base"
+  else
+    fail=$((fail+1))
+    fails+=("$base: expected=$expected actual=$actual")
+    printf '  FAIL  %-12s  %s  (got %s)\n' "$expected" "$base" "$actual"
+  fi
+done
+
+printf '\n%d pass, %d fail\n' "$pass" "$fail"
+if [ "$fail" -ne 0 ]; then
+  printf '\nfailing fixtures:\n'
+  printf '  - %s\n' "${fails[@]}"
+  exit 1
+fi
+exit 0