[bradygaster]

Closes #1842

The bug

The prompt-budget gate in test/gh-aw-quality.test.ts asserted a gh-aw prompt ceiling that does not exist, and measured a quantity gh-aw never sends. Both sides of the comparison were wrong, which left it reporting ~26 bytes of headroom and failing correct changes on byte count alone.

1. The ceiling is misattributed

The comment claimed "gh-aw enforces a hard 100 KB prompt ceiling (102 400 bytes)". Verified against gh-aw:

• 102400 is defaultRepoMemoryMaxFileSize in github/gh-aw → pkg/workflow/repo_memory.go — "the default maximum file size in bytes (100KB)" for repo-memory files. Unrelated to prompts.
• .github/aw/workflow-constraints.md never mentions a prompt ceiling.
• .github/aw/token-optimization.md treats prompt size purely as cost/quality ("Strip redundant instructions, examples, and pleasantries from the prompt body"), never as a byte limit.

2. It measured source, not the prompt

gh-aw strips every inline ## skill: block during the setup/interpolation step — per gh-aw .github/aw/skills.md:

Extraction runs in the setup/interpolation step (not at compile time): gh-aw writes each block to engine-specific skill locations and removes it from the main prompt body.

Measured composition of what the gate was summing:

	bytes
What the test measured (raw source sum)	97,254
What actually enters initial context (ambient)	31,855
Extracted to on-demand skills	65,398 (67%)

shared/squad-planning-ontology.md (16,420 B) and shared/squad-planning-policy.md (4,538 B) are entirely skill blocks — they start with ## skill: on line 1 and are fully extracted. workflows/squad.md is 69,608 B of which only ~25,167 B is prompt body; 21 skill blocks begin at L540.

3. It ignored what gh-aw does inject

Compiling for real (gh aw compile .github/workflows/squad.md --strict, exit 0) shows GH_AW_PROMPT_CONFIG carrying ~17 prompt items, including eight gh-aw boilerplate files the test counts at zero: xpia.md, temp_folder_prompt.md, markdown.md, safe_outputs_prompt.md, safe_outputs_create_pull_request.md, mcp_cli_tools_prompt.md, cli_proxy_with_safeoutputs_prompt.md, pr_context_prompt.md. The lock also defers our four files via {{#runtime-import}} (GH_AW_PROMPT_CONTENT_0005–0008), confirming they are assembled at runtime rather than inlined at compile time.

The fix

The correct budget already existed in this file. The gh-aw: inline skill extraction suite runs a faithful mirror of gh-aw's extractor (extractInlineSkills, itself modeled on setup/js/extract_inline_skills.cjs) and asserts keeps the ambient prompt under 40 KB against post-extraction ambient bytes. That is the real gate, and it passes at ~31.8 KB.

So this change does not add a second budget — it stops the redundant one from lying:

• Removed the fabricated GH_AW_PROMPT_CEILING_KB / 102400 constant and the false comment.
• Collapsed the two byte assertions into a single source-growth guard at 160 KB, explicitly documented as not a gh-aw limit, whose failure message points the reader at the ambient check first.
• Kept readText()'s CRLF normalization — it is why this math is platform-independent (.gitattributes pins eol=lf for many types but not *.md, so Windows worktrees would otherwise read +2,093 bytes).

Impact

Real headroom is ~70 KB, not 26 bytes. This unblocks #1801, #1780, #1730, #1757, and #1608, none of which needed to compact shell one-liners or delete rationale to buy bytes.

It also changes #1834: the shell-input security contract (workflows/squad.md L245–299, 2,695 B) can move into a ## skill: block rather than being deleted for budget reasons. #1834 is no longer a prerequisite for this issue.

Validation

npx vitest run test/gh-aw-quality.test.ts
Tests  91 passed | 13 skipped (104)

The 13 skips are pre-existing and tracked in #1833 (HAS_POSIX_SHELL gating at L30 skips two describe blocks on Windows).

Test-only change — no packages/*/src/ touched, so no changeset required.

[github-actions]

🛫 PR Readiness Check

ℹ️ This comment updates on each push. Last checked: commit 6a517bc

PR Scope: 🔧 Infrastructure

⚠️ 1 item(s) to address before review

Status	Check	Details
✅	Single commit	1 commit — clean history
✅	Not in draft	Ready for review
✅	Branch up to date	Up to date with dev
❌	Copilot review	No Copilot review yet — it may still be processing
✅	Changeset present	No source files changed — changeset not required
✅	Scope clean	No .squad/ or docs/proposals/ files
✅	No merge conflicts	No merge conflicts
✅	Copilot threads resolved	No Copilot review threads
✅	CI passing	All checks passing

Files Changed (1 file, +28 −21)

File	+/−
test/gh-aw-quality.test.ts	+28 −21

Total: +28 −21

---

This check runs automatically on every push. Fix any ❌ items and push again.
See CONTRIBUTING.md and PR Requirements for details.

[github-actions]

🟢 Impact Analysis — PR #1849

Risk tier: 🟢 LOW

📊 Summary

Metric	Count
Files changed	1
Files added	0
Files modified	1
Files deleted	0
Modules touched	1

🎯 Risk Factors

• 1 files changed (≤5 → LOW)
• 1 module(s) touched (≤1 → LOW)

📦 Modules Affected

tests (1 file)

• test/gh-aw-quality.test.ts

---

This report is generated automatically for every PR. See #733 for details.

[Copilot]

🟢 Approval recommended

The change is test-only and replaces an incorrect, misleading budget assertion with checks aligned to the workflow’s actual prompt assembly behavior.

Pull request overview

This PR fixes the gh-aw quality “prompt budget” gate to budget what is actually delivered (ambient prompt after inline skill extraction) rather than raw authored markdown bytes, and removes a misleading/incorrect 100KB “prompt ceiling” assertion.

Changes:

• Removed the fabricated 100KB prompt-ceiling constant and the headroom-based regression assertion tied to it.
• Consolidated the budget checks into the existing ambient-prompt (<40KB) gate plus a separate, explicit source-growth guard (160KB) that is documented as not a gh-aw limit.
• Updated test messaging/comments to explain the distinction between authored source size and delivered ambient prompt size.

File summaries

File	Description
test/gh-aw-quality.test.ts	Removes incorrect prompt-ceiling/headroom checks and replaces them with a documented source-growth guard while relying on the canonical ambient prompt budget.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.