Section 03: SSOT — compose-intel-summary helper

Status: Complete (2026-04-14) — all subsections complete, 3-round TPR converged clean, impl-hygiene review clean, section-close sweep identified 1 cross-subsection tooling opportunity tracked in §07.4. Goal: Establish .claude/skills/dual-tpr/compose-intel-summary.md as the ONE canonical source for the intel-pre-query template. Every current consumer (18 files: 6 review-family + 12 wider-skill consumers) @-includes it. Additional migrations are planned per §05 (/verify-tpr, /sync-claude, /fix-next-bug — these skills will add intel reconnaissance) and §07 (.claude/hooks/pre-review-intel.sh — a hook-family consumer that does not yet exist). The harness splices the included file into the prompt at expansion time — so updates to the template automatically propagate to all consumers without drift.

Context: The 2026-04-14 TPR (codex finding TPR-XX-007, severity high, category LEAK:algorithmic-duplication) confirmed that the availability-check → file-symbols → callers → callees → similar → bounded-summary pattern is inlined in 6 separate files today. This is a textbook SSOT violation per .claude/rules/impl-hygiene.md §Algorithmic DRY. The existing polling-protocol.md in the same directory is the canonical precedent — it consolidated three near-identical inlined polling blocks into one @-include after a similar drift-driven surfacing in early April 2026. This section repeats that consolidation for the intel-pre-query pattern.

Reference implementations:

Ori .claude/skills/dual-tpr/polling-protocol.md: exact SSOT pattern — one canonical file, many @-include consumers, the rule file (intelligence.md or impl-hygiene.md) explains the drift violation resolved
Ori .claude/skills/dual-tpr/compose-rules-brief.md: sibling in the same directory; shows the Agent(subagent_type=Sonnet) composition style that intel-summary may also adopt for longer summaries

Depends on: Section 01 (Skill dir structure for .claude/skills/dual-tpr/ must be stable; §01’s SKILL promotion work happens under .claude/skills/query-intel/, not here, but Skill-dir conventions must be settled before §03 adds a new sibling).

03.1 Write compose-intel-summary.md

File(s): .claude/skills/dual-tpr/compose-intel-summary.md (new)

The helper is a reference document, not an executable script. Consumers @-include it into their prompts. Its body is the canonical instructions that Claude (or a sub-agent) follows to produce a bounded Intelligence Summary.

Create .claude/skills/dual-tpr/compose-intel-summary.md with body:

# Intelligence Summary Injection — Canonical SSOT

**Single source of truth** for pre-query / summary-injection behavior across
ALL consumers: /tpr-review Step 0.75, /review-work Step 1.5, /review-plan,
/independent-review, /review-bugs, /tp-help, /verify-tpr, /sync-claude,
/fix-next-bug, /fix-bug, and the .claude/hooks/pre-review-intel.sh hook.
Every such consumer MUST @-include this file from its intel section rather
than inlining the pattern.

## Protocol

### Step A — Availability check

Bash (foreground): scripts/intel-query.sh status


Parse the JSON. If `status != "ok"`, skip silently. Do NOT emit an empty
section in the consumer's prompt — skipping means NO Intelligence Summary
block appears at all.

### Step B — Subsystem and symbol identification

For code/plan review modes, use the same scope the consumer is operating
on (e.g., `git diff --name-only HEAD~5..HEAD`). For custom-objective mode,
extract relevant file paths or symbol names from the objective text.

Map subsystems to presets per `.claude/rules/intelligence.md` §Subsystem
Mapping. DO NOT hardcode the mapping here.

### Step C — Run the queries

Up to 5 queries total to keep the summary bounded. Output is visible in
Claude's context — do NOT capture into a variable.

1. Subsystem preset OR directed search:
   `scripts/intel-query.sh --human <preset-or-search> --limit 5`
2. For top 3-5 changed files:
   `scripts/intel-query.sh --human file-symbols "<path-fragment>" --repo ori`
3. For each high-signal symbol:
   `scripts/intel-query.sh --human callers "<symbol>" --repo ori`
   `scripts/intel-query.sh --human callees "<symbol>" --repo ori`
   `scripts/intel-query.sh --human similar "<symbol>" --repo rust,swift,go --limit 5`

If any query returns empty, skip it silently in the summary.

### Step D — Condense into a bounded Intelligence Summary (≤500 chars)

Format:

Intelligence Summary (from intelligence graph):

[rust#12345] Similar bug / pattern — short phrase (N reactions)
[swift#6789] Reference implementation — short phrase
[ori] called by N sites across M modules — blast radius note


Rules:
- Maximum 5 bullets.
- Maximum 500 characters (hard cap; truncate with `…` if needed).
- Reference-repo citations use `[repo#N]` issue shorthand or
  `[repo:path]` for symbol results.
- Ori citations use `[ori]` prefix.
- Do NOT cite a result as authoritative — this is DISCOVERY for the
  consumer, not conclusions.

### Step E — Inject into the consumer's prompt

The consumer is responsible for placing the summary into its own prompt
template (e.g., after the `## Scope:` header in a reviewer prompt,
after the objective in a custom-objective prompt). This helper produces
the summary text; the consumer chooses where to place it.

## Graceful degradation

If `scripts/intel-query.sh status` returns unavailable, the entire
summary is OMITTED. Do NOT emit an empty "Intelligence Summary: no
results" block — that's noise. The consumer's prompt should be
syntactically valid whether or not the summary appears.

## Banned patterns

- Inlining this template in any consumer instead of `@`-including it
- Open-coding Neo4j access (bypassing `scripts/intel-query.sh`)
- Emitting a summary without the availability check
- Citing a graph result without verifying against actual code

## Consumers

Every consumer of this file references it via `@.claude/skills/dual-tpr/compose-intel-summary.md`
at its intel section. Updates to this protocol propagate automatically.

## Related

- `.claude/rules/intelligence.md` — when-to-query workflow inventory, subsystem mapping
- `.claude/skills/query-intel/SKILL.md` — full capability surface
- `scripts/intel-query.sh` — the canonical wrapper (206 lines; see §08 for planned UX improvements)
- `.claude/skills/dual-tpr/polling-protocol.md` — sibling SSOT for dual-source polling
- `.claude/skills/dual-tpr/compose-rules-brief.md` — sibling SSOT for rules-brief composition

Verify: wc -l .claude/skills/dual-tpr/compose-intel-summary.md reports ≥80 lines; file parses as valid markdown.
Subsection close-out (03.1):
- SSOT file created and content-reviewed for completeness (100 lines)
- Update 03.1 status to complete
- Run /improve-tooling retrospectively on 03.1 — Retrospective: no tooling gaps. The polling-protocol.md sibling was a perfect template reference; writing the new SSOT was a 1-step Write operation using the verbatim body specified at plan L62-162. No query patterns surfaced during composition that aren’t already in the template (availability check → file-symbols → callers/callees → similar → bounded summary). Sibling-precedent pattern (polling-protocol.md + compose-rules-brief.md) is itself the tool — it makes subsequent SSOTs near-mechanical.
- Run /sync-claude on 03.1 — .claude/rules/intelligence.md cross-reference deferred to §03.3 Completion Checklist, where the §Algorithmic DRY entry is bundled with the impl-hygiene.md update. No new rules-level change needed solely from §03.1; the SSOT is the product, the cross-reference lands after invariant verification.
- Repo hygiene check — clean (only the new SSOT file added; no temp files).

03.2 Replace inlined copies with @-includes

File(s): 18 consumer files migrated (6 review-family + 12 wider-skill consumers — scope expanded during execution)

The replacement is mechanical but order matters: verify each file’s inlined block matches the canonical template semantically before replacing, so we don’t lose domain-specific customization.

Scope discovery (2026-04-14): The initial plan targeted 6 review-family files identified by the TPR. During migration, grep -l 'scripts/intel-query.sh status' .claude/ -r surfaced 12 ADDITIONAL skills that also inline the pattern. Per CLAUDE.md zero-deferral + correctness invariants, all 18 were migrated in this subsection. The SSOT’s Step F was extended to document the full set of domain-specific extensions.

Target files (review-family — original 6 from TPR):

.claude/skills/review-work/SKILL.md:251-259 (Step 1.5 CONDITIONAL — Intelligence Pre-Query)
.claude/skills/tpr-review/SKILL.md Step 0.75 (~42 lines)
.claude/skills/review-plan/SKILL.md and its step-*.md protocol files (the §03 migration applies to the skill’s @-include surface — primarily step-5-editor.md’s “Intelligence Reconnaissance” block; /review-plan is implemented as a multi-step skill with no command-file surface)
.claude/commands/review-work.md:70-75
.claude/commands/independent-review.md:221-224
.claude/commands/review-bugs.md:152-178

Target files (wider skills — 12 discovered during execution):

.claude/skills/tp-help/SKILL.md:98 (inline mention)
.claude/skills/add-bug/SKILL.md:90 (inline mention)
.claude/skills/improve-tooling/SKILL.md:89 (inline mention)
.claude/skills/design-pattern-review/SKILL.md:134-155 (STEP 1.5)
.claude/skills/create-draft-proposal/SKILL.md:61-75 (Step 4.5)
.claude/skills/fix-bug/SKILL.md:105-120 (5a Intelligence Graph Query)
.claude/skills/impl-hygiene-review/SKILL.md:191-196 (Intelligence-assisted map)
.claude/skills/review-draft-proposal/SKILL.md:98-103 (CONDITIONAL Prior Art)
.claude/skills/create-plan/SKILL.md:377-382 (Step 2.5)
.claude/skills/rosetta-test/SKILL.md:75-80 (I. Cross-Language Intelligence)
.claude/skills/code-journey/SKILL.md:112-117 (Intelligence map)
.claude/skills/continue-roadmap/SKILL.md:381-399 (Step 2.1)

03.3 Verify SSOT invariant

File(s): N/A (verification-only)

Run the invariant check:

# Count inlined copies of the pre-query pattern (must be limited to the SSOT + legitimate teaching surfaces)
grep -l 'scripts/intel-query.sh status' .claude/ -r
# Result: 3 files — .claude/skills/dual-tpr/compose-intel-summary.md (SSOT itself),
#                   .claude/rules/intelligence.md (canonical when-to-query rule),
#                   .claude/commands/query-intel.md (command wrapper)
# All skill/command consumers migrated. Invariant satisfied.

Harness-side check: the @-include expansion is the harness’s standard splicing behavior (used by existing polling-protocol.md + compose-rules-brief.md SSOTs in the same directory). No additional harness-specific verification needed — the pattern is proven.
Document the SSOT invariant in .claude/rules/impl-hygiene.md — added “Precedents — SSOT-via-@-include for Skill Protocols” subsection to §Algorithmic DRY, citing both polling-protocol.md (2026-04) and compose-intel-summary.md (2026-04-14) precedents.
Subsection close-out (03.3):
- Invariant check passes (SSOT + 2 legitimate teaching surfaces; all consumer skills/commands migrated)
- Update 03.3 status to complete
- Run /improve-tooling retrospectively on 03.3 — Retrospective: the grep invariant is the only verification needed for now; promoting it to a pre-commit hook would be premature (the hook surface is already busy, and grep -l is fast enough that a periodic manual check suffices). When §07 ships .claude/hooks/pre-review-intel.sh, revisit — a UserPromptSubmit hook could run the invariant check alongside its primary intel pre-query work. Filed as a note for §07 to consider.
- Run /sync-claude on 03.3 — .claude/rules/impl-hygiene.md §Algorithmic DRY updated with Precedents subsection citing both dual-tpr SSOTs. Commit bundled with §03.3 closure.
- Repo hygiene check — clean (no temp files).

03.R Third Party Review Findings

Dual-source TPR run on 2026-04-14 (scratch dir /tmp/ori-tpr-a4vtapRS): codex walltime 547s / gemini walltime 176s. ASYMMETRY: HIGH but both reviewers independently identified the same core DRIFT class (SSOT Step F vs actual consumer usage) — agreement in spirit, not in exact (location, title). Codex found 3 additional findings gemini missed. All findings verified against the actual code before acceptance.

[TPR-03-001-codex][medium] .claude/skills/rosetta-test/SKILL.md:75 — Convert the four inline @-include mentions into real standalone splice sites. Resolved: Fixed on 2026-04-14. All four files (tp-help/SKILL.md:98, add-bug/SKILL.md:90, improve-tooling/SKILL.md:89, rosetta-test/SKILL.md:75) rewritten so the @-include appears on its own line with surrounding local context before/after. Invariant grep -c '^[[:space:]]*@\.claude/skills/dual-tpr/compose-intel-summary\.md$' .claude/ -r now returns 18 files (matching the 18 consumer count exactly). Evidence: tp-help/SKILL.md:98, add-bug/SKILL.md:90, improve-tooling/SKILL.md:89, and rosetta-test/SKILL.md:75 embed @.claude/skills/dual-tpr/compose-intel-summary.md mid-sentence inside prose/bullets rather than on their own line. The other 14 consumer sites use standalone @-include directives. polling-protocol.md:108 documents @-include as a full-content splice; mid-sentence placements are not valid include boundaries. Impact: GAP:ssot-splice-point — if the documented splice semantics apply strictly, these four consumers either expand into malformed prompt text or stop being true SSOT consumers, overstating the 18-consumer migration claim. Required plan update: Convert each of the four inline mentions into a standalone @-include line with surrounding local context before and after. Basis: inference. Confidence: medium.
[TPR-03-002-codex][medium] .claude/skills/dual-tpr/compose-intel-summary.md:4 — Restrict the SSOT consumer roster to actual current adopters. Resolved: Fixed on 2026-04-14. SSOT header rewritten to enumerate the 18 current consumers (6 review-family + 12 wider-skill) explicitly and move the future-intended consumers (/verify-tpr, /sync-claude, /fix-next-bug, .claude/hooks/pre-review-intel.sh) into a distinct “Planned future consumers” section with pointers to §05 and §07 where their migrations are tracked. Evidence: The SSOT header claims coverage over /verify-tpr, /sync-claude, /fix-next-bug, and .claude/hooks/pre-review-intel.sh. Verified: verify-tpr/SKILL.md, sync-claude/SKILL.md, fix-next-bug/SKILL.md contain zero compose-intel-summary references, and the hook path does not exist in the repo. Impact: DRIFT:ssot-consumer-inventory — the canonical file overstates adoption, inviting LEAK:scattered-knowledge when readers assume nonexistent adopters are already migrated. Required plan update: Trim the SSOT Consumers section to the actual current adopters (18 files, enumerated). Move /verify-tpr, /sync-claude, /fix-next-bug, and the hook into a distinct “Future intended consumers” list with pointers to §05/§07 where their migrations are planned. Basis: direct_file_inspection. Confidence: high.
[TPR-03-003-codex][low] .claude/skills/dual-tpr/compose-intel-summary.md:75 — Synchronize Step F with the migrated consumer extensions. Resolved: Fixed on 2026-04-14. Step F rewritten as the canonical “Consumer extension registry” with one entry per consumer, each listing the consumer’s actual scripts/intel-query.sh invocations verbatim. Added /design-pattern-review entry (previously missing). Corrected /fix-bug entry (added search/fixed). Corrected /rosetta-test entry (added file-symbols/search). Added a Registry Contract paragraph stating that any consumer adding/removing queries MUST update Step F in the same commit — codifying this as an SSOT obligation to prevent future DRIFT:intel-extension-registry. Evidence: Step F says /fix-bug extends the base protocol with callers + similar only, but the migrated fix-bug/SKILL.md:111-114 uses search + fixed + similar. Step F has no /design-pattern-review entry even though that consumer’s text explicitly labels its queries “per SSOT Step F”. The shared code-journey/rosetta-test bullet omits rosetta-test’s search "<failure mode>" query. [TPR-03-001-gemini] surfaced the same drift class independently. Impact: DRIFT:intel-extension-registry — Step F is no longer a faithful registry of consumer-specific deviations, so future migrations could silently lose behavior or copy the wrong extension set. Required plan update: Rewrite Step F as a maintained registry, reading each consumer’s actual queries directly and recording them verbatim. Add missing entries (/design-pattern-review) and correct incomplete entries (/fix-bug, /rosetta-test). Basis: git_history. Confidence: high.
[TPR-03-004-codex][medium] plans/query-intel-adoption/section-03-compose-intel-summary-ssot.md:4 — Refresh section 03 status and success criteria after the 18-consumer expansion. Resolved: Fixed on 2026-04-14. Frontmatter status → in-progress (§03.R/§03.N still pending). goal rewritten to reflect 18-consumer scope. success_criteria updated with true consumer count, new 3-file invariant (SSOT + 2 teaching surfaces), and Step F registry mention. Body intro updated to “Status: In Progress (§03.1 / §03.2 / §03.3 complete; §03.R triage in progress; §03.N pending)” with accurate file counts. Completion Checklist migrated from “6 consumers” to “18 consumers” with correct invariant; Exit Criteria rewritten to reflect the 3-file outcome. Evidence: Frontmatter and body still say status: not-started and “6 consumers” (success_criteria line 9, body line 40-41). §03.2 and §03.3 record 18 migrated consumers and a 3-file intel-query.sh status invariant. The Completion Checklist and Exit Criteria still reference the obsolete 6-consumer / exactly-one-file targets. Impact: DRIFT:plan-state and DRIFT:mission-criteria — /continue-roadmap and human readers will see the section as unstarted and scoped to the old 6-consumer target. Required plan update: Update frontmatter status → in-progress (§03.R/§03.N still pending), success_criteria to “18 consumers”, body intro to reflect true scope, Completion Checklist consumer count, and Exit Criteria invariant text. Basis: direct_file_inspection. Confidence: high.
[TPR-03-001-gemini][medium] .claude/skills/dual-tpr/compose-intel-summary.md:88 — Sync Step F domain extensions with actual consumer usage. (Semantic agreement with [TPR-03-003-codex] — different location/title, same root cause.) Resolved: Fixed on 2026-04-14. Same fix as [TPR-03-003-codex]: Step F rewritten as a maintained registry matching each consumer’s verbatim query set. Evidence: Step F lists /review-bugs using search + fixed, but actual usage is search + fixed + callers + file-symbols + similar. /rosetta-test lists symbols + callers/callees + similar, actual usage adds file-symbols + search. /fix-bug lists callers + similar, actual usage is search + fixed + similar. Impact: DRIFT — Step F documentation is out of sync with consumer code. Required plan update: Rewrite Step F to accurately list the exact queries used by each consumer extension (bundled with [TPR-03-003-codex]). Basis: direct_file_inspection. Confidence: high.
[TPR-03-002-gemini][informational] — Positive confirmation of invariant check (3 files contain the pattern: SSOT + 2 teaching surfaces). Non-actionable, no fix needed.

Round 2 findings (2026-04-14, `/tmp/ori-tpr-ZPNF2wWA`)

Re-review after fixes landed in commit dc1086ce. Dual-source run: codex walltime 308s (33 files read, thorough), gemini walltime 96s (15 files read — up from 8 in round 1, rules consulted CLAUDE.md + both rules, scope materially broader per the Thoroughness Re-review Directive). ASYMMETRY: HIGH but gemini’s floor is now adequate per §6b (rules complete, scope covers the meat of the change). Gemini emitted a clean-pass informational finding confirming fixes are sound. Codex caught 1 final low-severity DRIFT detail:

[TPR-03-005-codex][low] .claude/skills/dual-tpr/compose-intel-summary.md:132 — Align /create-plan Step F entry with consumer’s verbatim query set. Evidence: Step F /create-plan entry said “Then Step C base queries on high-signal symbols”, deferring to Step C’s default --repo rust,swift,go. But the consumer (create-plan/SKILL.md:383) uses similar "<symbol>" --repo rust,swift,go,koka --limit 5 — the koka override was dropped from Step F’s documentation. Impact: DRIFT:intel-extension-registry (micro) — Step F under-documented /create-plan’s actual prior-art surface. Maintainers reading Step F as the registry could silently lose the Koka leg of reconnaissance. Resolved: Fixed on 2026-04-14. Step F /create-plan entry now lists the callers/callees/similar queries verbatim with the --repo rust,swift,go,koka override explicitly noted, plus a parenthetical explaining why koka is included (plan reconnaissance benefits from Koka’s effect-system prior art). Basis: direct_file_inspection. Confidence: high.
[TPR-03-003-gemini][informational] — Round 2 clean-pass confirmation. Gemini verified all 5 round-1 fixes hold, confirmed the 3-file / 18-file invariants via shell queries, and recommended “§03 may proceed to closure.” Non-actionable.

Round 3 findings (2026-04-14, `/tmp/ori-tpr-cFZdEzZU`) — CLEAN PASS

Final re-review after round 2 fix landed in commit a54af6b6. Dual-source run: codex walltime 127s (7 files read, 3 rules, 5 diagnostics, basis: fresh_verification), gemini walltime 70s (confidence 100, needs_escalation: false). ASYMMETRY: MODERATE — acceptable for the narrow round-2-only scope. Merge summary: 0 actionable findings, 2 informational positive confirmations (one from each reviewer). Both reviewers independently confirm: /create-plan Step F entry now documents the koka override verbatim, consumer matches, invariants hold (3 teaching-surface files + 18 standalone @-include splice lines). TPR loop complete — both reviewers report full consensus that §03 is architecturally clean.

[TPR-03-006-codex][informational] — Round 3 clean-pass confirmation. Codex verified the round-2 Step F /create-plan fix is isolated and invariants still hold; “On the SSOT/consumer/invariant axis, §03 is clean; remaining work is close-out bookkeeping rather than another repair cycle.” basis: fresh_verification. Non-actionable.
[TPR-03-004-gemini][informational] — Round 3 clean-pass confirmation. Gemini verified the /create-plan Step F entry matches consumer perfectly. Invariant counts match (3 files / 18 splice lines). “plans/query-intel-adoption is fully clean.” Non-actionable.

TPR verdict: Full consensus clean across 2 rounds of fixing + 1 final confirmation round. third_party_review.status → resolved.

03.N Completion Checklist

Exit Criteria: Exactly 3 files under .claude/ contain the scripts/intel-query.sh status pattern — the SSOT itself (.claude/skills/dual-tpr/compose-intel-summary.md) plus two legitimate teaching surfaces (.claude/rules/intelligence.md canonical when-to-query rule and .claude/commands/query-intel.md command wrapper). All 18 current consumers reference the SSOT via standalone @-include splice lines. Step F registry in the SSOT matches each consumer’s actual query set verbatim (registry contract). .claude/rules/impl-hygiene.md §Algorithmic DRY §Precedents subsection cites the consolidation as a second SSOT-via-@-include precedent alongside polling-protocol.md. ./test-all.sh green modulo pre-existing unrelated failures.