Section 03: Orchestrator Rewrite

Status: Not Started Goal: Make the rewrite a localized SKILL.md edit that swaps Agent dispatch for Bash dispatch, lifts the bounded retry into the orchestrator, and trims the §9 / §11.P prose accordingly. After this section: §5 round-loop logic, §8a scratch-dir setup, §8b prompt composition, §8e shadow-edit absorption, §10 plan-TPR integration, §11 round summary, and the §11.5 autonomous carve-out are all unchanged. The rewrite is purely transport-layer.

Success Criteria: (echoes frontmatter — verifiable from the diff against tpr-review/SKILL.md HEAD)

§8c dispatch shape: len(REVIEWERS) foreground Bash tool calls in one assistant message, no Agent({subagent_type:..., model:'sonnet'}) calls
§8c.1 (NEW) extraction shape: len(REVIEWERS) Bash extract-report.py + len(REVIEWERS) Read calls, both batches parallel within their respective assistant messages
§8c.2 (NEW) retry shape: per-reviewer one-shot retry on tier 4.5/5 first-attempt failures, gated to ≤1 retry per reviewer per round
§9 table: 3-status surface preserved; defensive “sub-agent contract violation” row removed; new bash-failure triggers documented
§11.P banner: “Sonnet sub-agents” → “Bash dispatches”; wall-clock envelope preserved verbatim
Reviewer-set SSOT (I28) preserved by inspection of the dispatch pseudocode

Context: Per §01.2’s resolutions, all four open design questions resolved in favor of the most direct rewrite: harness’s native parallel-tool-call concurrency replaces the Agent transport; orchestrator absorbs the one-retry gate; Skill() callers are unaffected; Monitor was already excised. This section turns that design surface into a concrete edit.

Reference implementations:

tpr-review/SKILL.md §8c (current Agent template, lines 624–767) — replaced by the Bash template below
tpr-review/SKILL.md §9 (current transport-shape table) — defensive row dropped
tp_agent_prompt.md §Step 3.5 — retry logic to lift (deleted in §04 along with the rest of the file)

Depends on: §01 (design resolutions), §02 (baseline measurement).

Intelligence Reconnaissance

Queries run 2026-04-25:

scripts/intel-query.sh status — graph available (Neo4j 5.26.24); recorded as proof of protocol.
Surface-applicability note: edits are confined to .claude/skills/tpr-review/SKILL.md (markdown). The Ori code-symbol index covers Rust crates only — no graph-resident symbol participates in this rewrite. Recording the unavailability-of-applicability as freeform prose per schema’s Recon-block contract.
Surrogate query (grep): grep -n 'subagent_type\|Agent({' .claude/skills/tpr-review/SKILL.md enumerates every Agent dispatch site before editing.

Results summary (≤500 chars) [repo:.claude/skills/tpr-review/SKILL.md]: Edits localized to ~80 lines (§8c rewrite, §9 table trim, §11.P banner one-word swap, §1 description rewrite). Reviewer-set SSOT (I28) drives dispatch-loop shape — rewrite preserves “iterate REVIEWERS, emit len(REVIEWERS) tool calls, PRE-FLIGHT GATE checks count” exactly; only the tool kind switches Agent→Bash. No [ori] graph-resident Rust symbol touches this surface; correctness verified by §04.3 corpus replay [repo:.claude/skills/improve-tooling/tpr-review-design.md].

03.1 Replace §8c Agent dispatch with parallel foreground Bash

File(s): .claude/skills/tpr-review/SKILL.md (§8c body)

The dispatch message shape is preserved bit-for-bit; only the tool kind changes. The PRE-FLIGHT GATE language stays — the gate now counts Bash calls instead of Agent calls.

Edit §8c step 8c — Replace the canonical template. Replace today’s three Agent({subagent_type:..., model:'sonnet', description:..., prompt: <tp_agent_prompt.md...>}) blocks with three Bash({command:'bash invoke-{R}.sh "$scratch"', timeout: 2700000}) blocks. Use a for R in REVIEWERS loop in the pseudocode rather than three named blocks — preserves I28.

Target shape (paste into SKILL.md):

<assistant message begin>
  <§11.P status text — ≤2 lines, mandatory>
  Bash({ command: 'bash .claude/skills/tpr-review/invoke-${REVIEWERS[0]}.sh "$scratch"',  timeout: 2700000 })
  Bash({ command: 'bash .claude/skills/tpr-review/invoke-${REVIEWERS[1]}.sh "$scratch"',  timeout: 2700000 })
  Bash({ command: 'bash .claude/skills/tpr-review/invoke-${REVIEWERS[N-1]}.sh "$scratch"', timeout: 2700000 })
<assistant message end>

Edit §8c PRE-FLIGHT GATE — Update tool-name and count rule. Today reads “ALL Agent() calls for this round are in the SAME assistant message”; update to “ALL Bash() dispatch calls for this round are in the SAME assistant message” + “Required Bash count in the dispatch message: EXACTLY len(REVIEWERS) Bash() calls”. Keep the dispatch-shape diagram, dispatch_violation reporting, and the survivor-mode-retry exception clause — they are tool-kind-agnostic.
Strip sub-agent prose from §8c. Today’s §8c says “Sub-agents are thin CLI transports with split concerns” + describes how the sub-agent runs bash invoke-{REVIEWER}.sh + describes Tier 0 / 1 / 2 / 3 / 4 / 4.5 / 5 extraction inside the sub-agent + describes I22 banned final-message shapes. Replace with a single concrete paragraph: “The orchestrator dispatches len(REVIEWERS) foreground Bash calls in one assistant message; each call invokes the per-reviewer wrapper invoke-{R}.sh which runs the CLI and tees stdout to $scratch/{R}-stdout.txt. Wall-clock per round is max({R0}, {R1}, {RN-1}) — the harness runs parallel tool calls concurrently. After all Bash calls return, see step 8c.1 for extraction.”
Subsection close-out (03.1) — MANDATORY before starting 03.2:
- §8c body edited with Bash-template + updated PRE-FLIGHT GATE + sub-agent prose stripped
- Update this subsection’s status in section frontmatter to complete
- Repo hygiene check — compiler_repo/diagnostics/repo-hygiene.sh --check clean

03.2 Insert post-dispatch extract-report.py + Read sequence

File(s): .claude/skills/tpr-review/SKILL.md (NEW §8c.1 inserted between §8c and §8d)

The extraction step lives entirely in the orchestrator now.

Insert §8c.1 — Extract per-reviewer reports. After the dispatch message returns (all Bash calls terminated with exit codes), the orchestrator emits ONE assistant message with len(REVIEWERS) parallel Bash calls invoking extract-report.py:

<assistant message begin>
  Bash({ command: 'python3 .claude/skills/tpr-review/extract-report.py "$scratch" --reviewer ${REVIEWERS[0]}'  })
  Bash({ command: 'python3 .claude/skills/tpr-review/extract-report.py "$scratch" --reviewer ${REVIEWERS[1]}'  })
  Bash({ command: 'python3 .claude/skills/tpr-review/extract-report.py "$scratch" --reviewer ${REVIEWERS[N-1]}' })
<assistant message end>

Each call’s stdout is one JSON line: {"reviewer": …, "tier": …, "status": …, "report_bytes": …, …}. Capture and parse per reviewer.

Insert §8c.1 — Read per-reviewer reports. After the extract-report.py batch returns, ONE assistant message with len(REVIEWERS) parallel Read calls:
```
<assistant message begin>
  Read({ file_path: '$scratch/${REVIEWERS[0]}-report.txt'  })
  Read({ file_path: '$scratch/${REVIEWERS[1]}-report.txt'  })
  Read({ file_path: '$scratch/${REVIEWERS[N-1]}-report.txt' })
<assistant message end>
```
Each report is bounded at 16 KB by extract-report.py’s _enforce_report_cap. Reports map directly into the existing parse_tpr_report() consumers in §5 round-loop pseudocode.
Update §5 round-loop pseudocode. The line [codex_out, gemini_out, opencode_out] = dispatch_parallel_thin_transports($scratch) (lines ~278 + ~169 in current SKILL.md) is replaced with a 3-step inline expansion: dispatch_parallel_bash($scratch) → extract_per_reviewer($scratch) → read_per_reviewer_report($scratch). Each step corresponds to one assistant-message round-trip per the §8c / §8c.1 templates above.
Subsection close-out (03.2) — MANDATORY before starting 03.3:
- §8c.1 inserted with both extraction and read templates, parameterized by REVIEWERS
- §5 round-loop pseudocode updated to call the three new helpers
- Update this subsection’s status in section frontmatter to complete
- Repo hygiene check — clean

03.3 Lift Tier 4.5/5 retry into the orchestrator

File(s): .claude/skills/tpr-review/SKILL.md (NEW §8c.2 inserted between §8c.1 and §8d)

Per §01.2 Q1 resolution: keep one bounded retry, lifted from sub-agent into orchestrator.

Insert §8c.2 — Bounded one-retry on Tier 4.5/5. After §8c.1’s extract step has produced JSON-line outputs for each reviewer, inspect the tier field. For each reviewer where tier ∈ {4.5, 5} on first attempt:
- Preserve the first attempt: cp $scratch/{R}-stdout.txt $scratch/{R}-stdout-attempt1.txt; cp $scratch/{R}-report.txt $scratch/{R}-report-attempt1.txt
- Re-invoke the wrapper (single foreground Bash, timeout: 2700000): bash invoke-{R}.sh "$scratch"
- Re-run extract-report.py "$scratch" --reviewer {R} against the fresh stdout
- Re-Read $scratch/{R}-report.txt
Selection rule (matches tp_agent_prompt.md §Step 3.5 selection logic verbatim):
1. If either attempt produced tier 1–4 (status: ok), use that attempt’s report. Both → use second.
2. Else if either attempt produced tier 4.5 (status: partial), use that attempt; both partial → merge per current rules (use attempt with more progress_notes: content; set summary: to “2 attempts, neither reached report emission”).
3. Else (both tier 5), emit status: failed with summary “2 attempts produced no extractable content”.
Pin the retry-budget invariant. Add a one-line invariant beneath the retry block: “MAX 1 retry per reviewer per round. Per-reviewer retries dispatch independently — a tier-4.5 retry on codex does NOT trigger re-dispatch of gemini/opencode.” This preserves the current sub-agent-internal retry semantics.
Document the retry as orchestrator-side in §8c.2 and remove the §Step 3.5 reference from §8c. Today §8c references “the sub-agent’s bounded internal retry per tp_agent_prompt.md §Step 3.5”; replace with “the orchestrator’s bounded one-retry gate per §8c.2”.
Subsection close-out (03.3) — MANDATORY before starting 03.4:
- §8c.2 inserted with retry trigger + selection rule + retry-budget invariant
- §8c sub-agent-Step-3.5 reference rewritten to point at §8c.2
- Update this subsection’s status in section frontmatter to complete
- Repo hygiene check — clean

File(s): .claude/skills/tpr-review/SKILL.md (§9 table; §11.P banner)

03.N Completion Checklist

Exit Criteria: grep -c 'Agent({.*sonnet' .claude/skills/tpr-review/SKILL.md returns 0 (or only matches a §5.5 /independent-review carve-out that explicitly survives); grep -n 'Bash({.*invoke-' .claude/skills/tpr-review/SKILL.md shows the per-reviewer dispatch sites; python -m scripts.plan_corpus check plans/inline-tpr-transport/section-03-orchestrator-rewrite.md exits 0.