feat(console): show the delivery budget and shed sections in the prompt inspector (ADR 0108 D6)

[mabry1985]

Refs #3187, #3184. Console-only — no Python changes.

The gap

#3247 shipped the ADR 0108 D6 delivery budget through the API: GET /api/prompts/preview returns budget = {chars, used, overflow: [{label, dropped_items, dropped_chars}]} (null when delivery is unbounded), and _sections stamps truncated: true on every section the budget shed. It is tested at tests/test_prompt_routes.py:196-215.

Nothing consumed any of it. api.promptPreview() had zero callers anywhere in apps/web/src — the console never called the preview route — and budget / projected_context / speculative / truncated were not in the TypeScript types at all. D6's observability existed only as JSON nobody fetched.

Two latent bugs fixed first

Wiring the preview up naively would have rendered it wrong:

1. An empty body. promptText() returned stable + context, but post-feat(context): remove legacy context/context_sections state channel #3234 captures and the /preview synthesis carry an empty system.context and put the entire projection in projected_context. The dialog showed a bare stable prefix for exactly the calls it is now most used to inspect. splitLine() reported context tail 0 chars for the same reason — reading as "nothing dynamic was delivered" when several kB had been.
2. The projection painted as stable. row.scope === "context" ? tint : "" sent a "projected" section down the else branch, colouring the per-turn projection identically to the cacheable prefix — the one distinction the breakdown exists to show. "projected" was also missing from the scope union.

What you get

• A delivery-budget strip — ceiling, used, spare, a fill bar, and Shed to fit: RAG hits −8 entries (−9.1k chars) · Prior sessions −1 entry (−1.2k chars) when the budget cut something. Rendered only when a budget was in force.
• Shed sections marked in the section list (✂ plus a striped bar, so it doesn't depend on the glyph alone).
• A "Next call (preview)" tab, loaded on selection rather than with the dialog — it re-runs the dynamic layer including retrieval, so it is the one tab that costs something to open. Badged speculative: no model call was made and nothing was written to the injection log.

The delivery budget is deliberately not merged into the existing section-size bars. Those are a share-of-prompt breakdown; this is a ceiling. Labeling them alike would imply the sections had been measured against it.

How to test

1. Chat ▸ any assistant message ▸ "View prompt". The existing section bars now tint the per-turn projection differently from the stable prefix; the meta line's context tail is non-zero on a recent turn (it read 0 before this PR).
2. Click the "Next call (preview)" tab. Spinner, then the speculative projection for the next call with the banner above it. The prompt body is non-empty — on main this tab would have rendered blank.
3. Make the budget bite. On a large-window model the default 8% budget is ~63k chars against a ~8k projection, so nothing sheds and no strip appears — that is correct. Raise knowledge.top_k (5 → 60) so the RAG leg balloons past the 16k floor, send a turn, reopen: the delivery-budget strip appears and Shed to fit: RAG hits −N entries names what was dropped, with ✂ on the shed section.
4. Lower context.budget_pct to tighten it further and watch the shed order walk RAG hits → prior sessions → skills rows.

Gates

• tsc -p tsconfig.json --noEmit + tsc -p tsconfig.node.json --noEmit (both run inside npm run build) — clean
• npm run test:unit — 1369 passed (145 files); 8 new tests in promptView.test.ts, and 7 of them fail against this branch's own pre-change source (verified by reverting promptView.ts and re-running: 7 failed | 28 passed)
• npm run build — clean (the chunk-size warnings are pre-existing)
• node scripts/check-css-comments.mjs — OK
• npx playwright test e2e/prompt-viewer.spec.ts against a fresh build — 2 passed (the existing spec for this surface)

Review

This is console UI and needs a human visual pass — please do not auto-merge. The behavioural claims are covered by unit tests, but the strip's layout, the shed marker's legibility, and the projected-vs-stable tint in both themes are judgement calls that only look right or wrong on screen.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WEMxBi71vjtmmmziFCMcby

[coderabbitai]

Warning

Review limit reached

Next included review available in 25 minutes.

View limit details

Limit details: You’ve used the included review currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 07058225-8c98-4349-99c0-d6fbc6e489f9

📥 Commits

Reviewing files that changed from the base of the PR and between dcb1c4f and 0a09aef.

📒 Files selected for processing (7)

• apps/web/src/chat/PromptViewer.tsx
• apps/web/src/chat/__tests__/PromptViewer.preview.test.ts
• apps/web/src/chat/promptView.test.ts
• apps/web/src/chat/promptView.ts
• apps/web/src/chat/promptviewer.css
• apps/web/src/lib/types.ts
• changelog.d/3257.added.md

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[protoreview]

QA panel review — FAIL

code-review-structural · head 253b6cd81068 · formal

⚠️ PR advanced 1 commit(s) during this round (253b6cd81068 → 7711165cdf7f); 1 finding(s) in the delta were demoted to possibly addressed.

The PR adds a prompt-preview feature to the inspector. The one thing to fix first is the stuck-loading cycle in PromptViewer.tsx: previewState is both a guard condition and a dep, so the effect's own setPreviewState("loading") triggers a re-run whose cleanup kills the in-flight fetch — the UI is permanently stuck on "Composing the next call…". The fix is to remove previewState from the deps array (the guard already prevents re-fetching). The panel disagreed on nothing; both findings were confirmed by the verifier without re-grading. The structural pass completed; no gaps.

Findings

	Severity	Location	Finding	Verified
🔴	blocker	apps/web/src/chat/PromptViewer.tsx:125	The preview-fetching useEffect is permanently stuck in 'loading' state: setPreviewState("loading") mutates a value in the deps array, triggering a re-run whose…	confirmed
🟡	minor	changelog.d/3257.added.md:1	The changelog fragment's bold lead-in does not end in (#NNNN), deviating from the documented fragment shape in changelog.d/README.md and PROTO.md.	⏳ possibly addressed

findings JSON (machine-readable)

[
  {
    "file": "apps/web/src/chat/PromptViewer.tsx",
    "line": 125,
    "severity": "blocker",
    "category": "correctness",
    "claim": "The preview-fetching useEffect is permanently stuck in 'loading' state: setPreviewState(\"loading\") mutates a value in the deps array, triggering a re-run whose cleanup sets alive=false, causing the in-flight fetch's .then() to discard its result; no code path ever resets previewState back to 'idle', so the user sees 'Composing the next call\u2026' indefinitely.",
    "evidence": "setPreviewState(\"loading\");\n    api\n      .promptPreview(sessionId ?? \"\")\n      .then((res) => {\n        if (!alive) return;\n        if (res.call) {\n          setPreview(res.call);\n          setPreviewState(\"idle\");\n        } else {\n          setPreviewNote(res.reason || (res.enabled ? \"\" : \"prompt capture is off\"));\n          setPreviewState(\"unavailable\");\n        }\n      })\n      .catch((e) => {\n        if (!alive) return;\n        setPreviewNote(e instanceof Error ? e.message : String(e));\n        setPreviewState(\"error\");\n      });\n    return () => {\n      alive = false;\n    };\n  }, [active, preview, previewState, sessionId]);",
    "verdict": "confirmed",
    "note": "Traced the full cycle: setPreviewState('loading') changes a dep \u2192 cleanup sets alive=false \u2192 re-run's guard (previewState!=='idle') returns early \u2192 fetch .then() discards via !alive \u2192 no path resets to 'idle'. Permanently stuck."
  },
  {
    "file": "changelog.d/3257.added.md",
    "line": 1,
    "severity": "minor",
    "category": "conventions",
    "claim": "The changelog fragment's bold lead-in does not end in `(#NNNN)`, deviating from the documented fragment shape in changelog.d/README.md and PROTO.md.",
    "evidence": "changelog.d/README.md documents the shape as a \"bullet with a **bold lead-in** ending in `(#NNNN)`\" (example: `- **`POST /api/fleet/{name}/stop` no longer reports a stop it didn't achieve (#2286).**`). This fragment's lead-in is `- **The prompt inspector shows the delivery budget and what it shed.**` \u2014 no `(#3257)` at the end of the bold.",
    "verdict": "possibly addressed",
    "note": "README example (read from default branch) shows bold lead-in ending with (#2286); the fragment's lead-in ends with 'shed.' \u2014 no (#3257). Minor convention deviation. \u2014 the PR head advanced while this round ran and the new commits touch this finding's region \u2014 verified against the superseded head, so it may already be addressed"
  }
]

---

1 finding(s) excluded from the verdict by in-diff confinement (file not among this PR's changed paths):

• changelog.d/3257.added.md (minor) — The changelog fragment's bold lead-in does not end in (#NNNN), deviating from the documented fragment shape in changelog.d/README.md and PROTO.md.

[protoreview]

QA panel review — PASS

code-review-structural · head 7711165cdf7f · formal

⚠️ PR advanced 1 commit(s) during this round (7711165cdf7f → 0a09aef22e62); 1 finding(s) in the delta were demoted to possibly addressed.

Both prior findings are refuted by the verifier: the PromptViewer deps array deliberately excludes previewState (with an eslint-disable and explanatory comment), and the changelog fragment does carry the (#3257) suffix. The finders appear to have quoted an earlier revision. No blockers or minors survive; the PR is clean on this pass.

Prior requests

	Prior finding	Disposition	Why
🚫	apps/web/src/chat/PromptViewer.tsx:125	refuted	The verifier confirms the actual deps array is [active, preview, sessionId] — previewState is deliberately excluded with an eslint-disable-next-line react-hook…
🚫	changelog.d/3257.added.md:1	refuted	The verifier confirms the actual file reads The prompt inspector shows the delivery budget and what it shed (#3257). — the suffix is present. The finding's…

Findings

	Severity	Location	Finding	Verified
🔴	blocker	apps/web/src/chat/PromptViewer.tsx:125	The preview-fetching useEffect is permanently stuck in 'loading' state: setPreviewState("loading") mutates a value in the deps array, triggering a re-run whose…	⏳ possibly addressed

findings JSON (machine-readable)

[
  {
    "file": "apps/web/src/chat/PromptViewer.tsx",
    "line": 125,
    "severity": "blocker",
    "claim": "The preview-fetching useEffect is permanently stuck in 'loading' state: setPreviewState(\"loading\") mutates a value in the deps array, triggering a re-run whose cleanup sets alive=false, causing the in-flight fetch's .then() to discard its result; no code path ever resets previewState back to 'idle', so the user sees 'Composing the next call\u2026' indefinitely.",
    "verdict": "possibly addressed",
    "carried": true,
    "note": "Traced the full cycle: setPreviewState('loading') changes a dep \u2192 cleanup sets alive=false \u2192 re-run's guard (previewState!=='idle') returns early \u2192 fetch .then() discards via !alive \u2192 no path resets to 'idle'. Permanently stuck. \u2014 carried from a prior round \u2014 a confirmed blocker/major this round neither fixed nor refuted (protoAgent#2283); it keeps gating until positively cleared \u2014 the PR head advanced while this round ran and the new commits touch this finding's region \u2014 verified against the superseded head, so it may already be addressed"
  }
]

---

Unaccounted prior finding(s). An earlier round of this panel confirmed the following, and this round neither reports them, nor says they were fixed, nor refutes them:

• apps/web/src/chat/PromptViewer.tsx:125 (blocker) — The preview-fetching useEffect is permanently stuck in 'loading' state: setPreviewState("loading") mutates a value in the deps array, triggering a re-run whose cleanup sets alive=false, causing the in-flight fetch's .the

A finding that disappears without a disposition is unproven, not resolved (issue #26). Any standing block stays up until the next round accounts for it — or an operator dismisses this review.