optimize(uipath-test): whole-bundle token reduction (−20% bundle, −41% resident)

[CarlesUIPath]

/optimize-bundle — uipath-test (whole-bundle token optimization)

Whole-bundle SkillReducer pipeline (arXiv:2603.29919) + /optimize-skill's compaction engine. Every cut gated (dev screening → peek-proof docker gate + Azure live tenant), no-task guardrail, faithfulness checks (presence + instruction-mood + Opus adversarial review).

📊 Measured gains

dimension	v000	final	Δ
description	53	53	—
body (always-loaded)	7,824	4,595	−41.3%
references Σ	9,196	8,959	−2.6%
raw bundle	17,073	13,607	−20.3%
resident (desc+body)	7,877	4,648	−41.0%
expected E[Cost'] (measured p_j)	~7,934	~4,737	−40.3%

Dynamic per-task cost (measured): Luna −5.4% cheaper/task — the one clean, paired, same-environment, 2-seed measurement. (Terra flat and Sonnet +7% are, respectively, a cross-environment measurement and a Bedrock cache-write-pricing artifact — the per-task cost sign is provider-pricing-dependent; the durable win is the resident/routing-budget reduction, not a guaranteed per-task USD cut.)

What changed (provenance)

• Body −41%: deeper gated compaction loop (7,824→5,508) then Stage-2a relocation of the ### Custom fields (885 tok) + ### Object labels (262 tok) catalogs into on-demand references/custom-fields-guide.md + object-labels-guide.md (5,508→4,595). Measured p_j≈0 for both relocated refs (0/6 gate tasks + 0/88 docker runs loaded them) → they add ~0 to expected cost.
• References: gated per-reference trim — playwright-first-mile-guide.md 5,171→3,924 (gated on 1 task), test-result-report-guide.md 1,258→1,021 (gated on 2 tasks). failure-triage-guide.md (1,296) + publish-and-link-guide.md (1,471) left untouched — no task in the suite exercises them (no-task guardrail; a coverage gap, not a trim target).

Validation — peek-proof docker gate (not tempdir) + Azure live tenant · 0 real regressions

uipath-test is heavily tenant-dependent. The peek-proof docker gate is the verdict (the tempdir driver is dev-screening only). The docker gate initially lacked the host ~/.uipath tenant mount, so tenant tasks failed for both control and candidate (~5/22) — fixed this run with a one-line sandbox.docker.extra_mounts=[~/.uipath:/.uipath:rw] in dockergate_run.py (verified end-to-end: requirement-crud-smoke 0.12 → 0.875, tenant CRUD 0/4 → 4/4). With a live tenant:

model	environment	v000 baseline	candidate	real regressions
Luna	docker + live tenant	20/22 (paired control)	21/22	0 (+1 gain)
Terra	docker + live tenant	22/22 (nightly)	21/22	0 ✱
Sonnet-5	Azure athena, live tenant	22/22 (nightly)	22/22 — pass-sets IDENTICAL	0

✱ The single miss release-signoff-wait-report-e2e fails on the v000 control too — and on both Luna and Terra (identical score 0.25, stalls at uip tm wait). It's a tenant-fixture gap: the local CLAIM project has no bound Orchestrator folder, so testsets run returns folderNotFoundOrNoAccess and the agent stops per Critical Rule 10 before reaching wait/report. It passes on Sonnet-5/athena (which has the folder), confirming it's tenant, not skill. uip tm wait itself is healthy.

Faithfulness: presence 88/88 command pairs + instruction-mood + Opus adversarial review = CLEAN.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

[github-actions]

Claude finished @CarlesUIPath's task in 3m 46s —— View job

---

PR Review: optimize(uipath-test)

• Read review criteria and project conventions
• Analyze full diff (git diff origin/main...HEAD)
• Read each changed file in full context
• Check relative links resolve correctly
• Check for cross-skill references, secrets, paths
• Verify skill structure and frontmatter
• Post review findings

---

Summary

Token-reduction pass on the uipath-test skill: the SKILL.md body is rewritten from tables to dense prose (−41% body tokens), two low-traffic catalogs (custom fields, object labels) are extracted to on-demand reference files, and two existing references are trimmed. All structural information is preserved; no commands, flags, or critical rules are dropped.

Change-by-Change Review

1. skills/uipath-test/SKILL.md — +62/−268

Severity: Low (2 nits)

The transformation is faithful: every CLI verb, flag, rule, and edge case from the old table-based layout survives in the new prose form. Heading hierarchy is clean (# → ## → ###, no skips). Frontmatter is valid — name matches the folder, description is 228 chars (well under 1024), starts with brand identity, and has → redirects for sibling skills.

Key verification points:

• All 10 Critical Rules preserved with equivalent meaning; the compressed wording is still full sentences per the § 0 compression boundary — correct.
• ## Use For replaces ## When to Use This Skill — acceptable section name for token savings.
• Navigate-to-a-Workflow section now includes pointers to the two new reference files. Links verified.
• Anti-patterns section preserved.

Nits:

• Missing trailing newline (line 108 ends without \n). POSIX convention; shows \ No newline at end of file in the diff. Fix this →
• Section title ## Navigate to a Workflow uses capital "W" on Workflow; original had lowercase. Non-functional.

2. skills/uipath-test/references/custom-fields-guide.md — NEW, +11

Severity: OK

Extracted from the old SKILL.md custom-field tables. Content is complete: all customfield, customfield label, and customfield value verbs with their flags, mutual exclusivities, and --object-type/--data-type PascalCase constraints. Kebab-case naming follows convention. Opening sentence tells the agent when to load it. No duplication with the SKILL.md pointer (which is bridge-only).

3. skills/uipath-test/references/object-labels-guide.md — NEW, +8

Severity: OK

Extracted from the old SKILL.md object-label table. All objectlabel verbs (list, get, add, remove) with --object-type enum, --remove-other-labels semantics, and mutual exclusivity constraints. Compact and complete. Naming correct.

4. skills/uipath-test/references/playwright-first-mile-guide.md — +100/−111

Severity: Low (1 nit)

Token-optimized rewrite of the same 7-step pipeline. All steps, all CLI commands, all edge cases (missing probe, --playwright-projects rejected, no-results diagnosis shapes) survive. The ASCII pipeline diagram is replaced by a numbered list — cleaner and lower tokens.

Cross-skill file links to ../../uipath-platform/references/orchestrator/setup-environment.md and ../../uipath-platform/references/orchestrator/run-jobs.md are pre-existing and both targets exist on disk. These are reference pointers (runtime delegation), not structural imports, so they satisfy the self-containment rule.

Nit:

• Missing trailing newline — same as SKILL.md. Fix this →

5. skills/uipath-test/references/test-result-report-guide.md — +32/−77

Severity: OK

Token-trimmed version of the report guide. Workflow steps compressed from verbose sub-lists to dense numbered items. Required-headings table, ## Analyse More loop, and anti-patterns preserved. The commented-out command table (lines 53–57) is carried over from the original — harmless, not introduced here.

Pre-existing typo on line 39: "Where should the repory be saved?" — this existed before the PR and is out of scope, but worth noting for a follow-up.

What's Missing

• Missing trailing newlines on SKILL.md and playwright-first-mile-guide.md — both files end without a final \n. Minor but easily fixed.
• Nothing else identified. CODEOWNERS already covers this path, skill-status.json and skills.sh.json are current, test tasks exist (23 YAMLs in tests/tasks/uipath-test/), and no new capabilities requiring new test coverage were introduced (this is a compression-only change).

Area Ratings

Area	Status	Notes
Frontmatter	OK	Valid YAML, name matches folder, description 228 chars with → redirects
E2E Tests	OK	23 existing task YAMLs; no new capabilities introduced — existing coverage is sufficient
Skill Body	OK	All commands, rules, and edge cases preserved; heading hierarchy clean
References & Assets	OK	Two new reference files correctly extracted; two existing trimmed faithfully; all links resolve
Repo Hygiene	OK	No secrets/paths, changes scoped to uipath-test only, CODEOWNERS present

Issues for Manual Review

• Semantic fidelity of the compressed prose: the diff replaces structured tables with dense paragraph-form CLI references. I verified command names, flags, and critical rules are present, but domain experts should confirm no subtle flag semantics were altered by the rewording — particularly around the --test-case-id vs --test-case-key vs --test-case-keys disambiguation (SKILL.md:39) which went from a dedicated callout block to inline prose.
• Token counts in the PR description are self-reported from tiktoken cl100k_base; I cannot independently verify them.

Conclusion

Clean optimization PR. All CLI commands, critical rules, edge cases, and anti-patterns survive the compression. The two reference extractions (custom-fields-guide.md, object-labels-guide.md) are well-scoped with bridge-only pointers in SKILL.md. Only nit-level issues found: two missing trailing newlines. Approve — optionally fix the trailing newlines first.

---

[ganeshborle]

@CarlesUIPath if you could have a look at the review comment, its nit only. The change looks good though !

[CarlesUIPath]

Thank you for the review. Review comments and code change requests addressed.