ci: harden npm publish workflow#7
Merged
Merged
Conversation
drewstone
added a commit
that referenced
this pull request
Apr 8, 2026
The competitive bench at v0.19.0 surfaced a real architectural bug in bad's planner: on extraction tasks the planner emits runScript→complete with placeholder values (null, "<from prior step>") in complete.result because the planner has to commit to result text BEFORE runScript runs. Pre-fix: dashboard-extract pass rate 0/3 = 0%. Three layers of defense: 1. executePlan placeholder substitution (deterministic, runner-side). Tracks lastRunScriptOutput across plan steps. When a complete.result matches hasPlaceholderPattern() AND there's a captured runScript output, substitutes the runScript output as the actual final result. Marks the substituted turn with "Gen 7.2 substituted" for forensics. 2. executePlan auto-complete-from-runScript. When the planner correctly emits ONLY runScript (no complete) per the new prompt rule, and the plan exhausts, the runner synthesizes a complete with the runScript output instead of falling through to the per-action loop. Eliminates 4-5 wasted per-action LLM calls. 3. Planner system prompt rule #7. New rule explicitly tells the planner: "EXTRACTION TASKS: when the goal asks you to READ, EXTRACT, REPORT, or RETURN values from the page, the LAST step of your plan MUST be runScript. Do NOT emit a complete step after the runScript with literal values, because at planning time you cannot know what runScript will return." Byte-stable so prompt cache still hits. hasPlaceholderPattern detects: - JSON null literals: {"x": null, "y": null} - Angle-bracket placeholders: <from prior step>, <placeholder>, <value from ...>, <extracted ...>, <observed ...>, <previous step> - Double-curly templates: {{userCount}} Conservative — "null" in prose like "null pointer exception" does NOT match because we look for the JSON null literal pattern (`: null`). Verified result (5 reps × dashboard-extract, isolated run, gpt-5.2): | metric | n | mean | min | max | |-----------------|---|--------:|--------:|--------:| | pass rate | 5 | 100% | - | - | | wall-time (s) | 5 | 7.7 | 5.1 | 9.4 | | turns | 5 | 2.0 | 2 | 2 | | LLM calls | 5 | 1.0 | 1 | 1 | | total tokens | 5 | 3835 | 3700 | 4015 | | cost ($) | 5 | 0.0131 | 0.0112 | 0.0156 | Wilson 95% CI on pass rate: [57%, 100%]. Pre-Gen 7.2 (v0.19.0): bad scored 0/3 = 0% on this task. Post-Gen 7.2 (this PR): 5/5 = 100% AND beats browser-use: | metric | bad (Gen 7.2) | browser-use 0.12.6 | verdict | |-----------|--------------:|-------------------:|------------------| | pass rate | 100% (5/5) | 100% (3/3) | tied | | wall-time | 7.7s | 20.6s | bad 2.7x faster | | LLM calls | 1.0 | 3.0 | bad 3x fewer | | tokens | 3835 | 19908 | bad 5.2x fewer | | cost | $0.0131 | $0.0258 | bad 49% cheaper | Tests: 944 passing (937 + 7 new): - placeholder substitution happy path - leave-unchanged when no placeholders - auto-complete-from-runScript when plan ends with successful runScript - does NOT auto-complete on empty runScript output - hasPlaceholderPattern: detects all 3 placeholder types, ignores prose and clean JSON Tier1 deterministic gate: PASSED (no regressions). Honest caveat: a concurrent 3-rep run during the full competitive grid (with parallel chromium contention from tier1-gate running) showed 2/3 = 67% — one rep had the LLM-generated runScript picking the wrong DOM element ("+12.5% from last month" subtitle instead of "12,847" value). That's LLM script quality, separate from this fix; tracked as a future Gen 7.3 follow-up. The Gen 7.2 mechanism is verified deterministic by the 7 unit tests.
7 tasks
drewstone
added a commit
that referenced
this pull request
Apr 8, 2026
…sks (#55) The competitive bench at v0.19.0 surfaced a real architectural bug in bad's planner: on extraction tasks the planner emits runScript→complete with placeholder values (null, "<from prior step>") in complete.result because the planner has to commit to result text BEFORE runScript runs. Pre-fix: dashboard-extract pass rate 0/3 = 0%. Three layers of defense: 1. executePlan placeholder substitution (deterministic, runner-side). Tracks lastRunScriptOutput across plan steps. When a complete.result matches hasPlaceholderPattern() AND there's a captured runScript output, substitutes the runScript output as the actual final result. Marks the substituted turn with "Gen 7.2 substituted" for forensics. 2. executePlan auto-complete-from-runScript. When the planner correctly emits ONLY runScript (no complete) per the new prompt rule, and the plan exhausts, the runner synthesizes a complete with the runScript output instead of falling through to the per-action loop. Eliminates 4-5 wasted per-action LLM calls. 3. Planner system prompt rule #7. New rule explicitly tells the planner: "EXTRACTION TASKS: when the goal asks you to READ, EXTRACT, REPORT, or RETURN values from the page, the LAST step of your plan MUST be runScript. Do NOT emit a complete step after the runScript with literal values, because at planning time you cannot know what runScript will return." Byte-stable so prompt cache still hits. hasPlaceholderPattern detects: - JSON null literals: {"x": null, "y": null} - Angle-bracket placeholders: <from prior step>, <placeholder>, <value from ...>, <extracted ...>, <observed ...>, <previous step> - Double-curly templates: {{userCount}} Conservative — "null" in prose like "null pointer exception" does NOT match because we look for the JSON null literal pattern (`: null`). Verified result (5 reps × dashboard-extract, isolated run, gpt-5.2): | metric | n | mean | min | max | |-----------------|---|--------:|--------:|--------:| | pass rate | 5 | 100% | - | - | | wall-time (s) | 5 | 7.7 | 5.1 | 9.4 | | turns | 5 | 2.0 | 2 | 2 | | LLM calls | 5 | 1.0 | 1 | 1 | | total tokens | 5 | 3835 | 3700 | 4015 | | cost ($) | 5 | 0.0131 | 0.0112 | 0.0156 | Wilson 95% CI on pass rate: [57%, 100%]. Pre-Gen 7.2 (v0.19.0): bad scored 0/3 = 0% on this task. Post-Gen 7.2 (this PR): 5/5 = 100% AND beats browser-use: | metric | bad (Gen 7.2) | browser-use 0.12.6 | verdict | |-----------|--------------:|-------------------:|------------------| | pass rate | 100% (5/5) | 100% (3/3) | tied | | wall-time | 7.7s | 20.6s | bad 2.7x faster | | LLM calls | 1.0 | 3.0 | bad 3x fewer | | tokens | 3835 | 19908 | bad 5.2x fewer | | cost | $0.0131 | $0.0258 | bad 49% cheaper | Tests: 944 passing (937 + 7 new): - placeholder substitution happy path - leave-unchanged when no placeholders - auto-complete-from-runScript when plan ends with successful runScript - does NOT auto-complete on empty runScript output - hasPlaceholderPattern: detects all 3 placeholder types, ignores prose and clean JSON Tier1 deterministic gate: PASSED (no regressions). Honest caveat: a concurrent 3-rep run during the full competitive grid (with parallel chromium contention from tier1-gate running) showed 2/3 = 67% — one rep had the LLM-generated runScript picking the wrong DOM element ("+12.5% from last month" subtitle instead of "12,847" value). That's LLM script quality, separate from this fix; tracked as a future Gen 7.3 follow-up. The Gen 7.2 mechanism is verified deterministic by the 7 unit tests.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
v*,agent-browser-driver-v*)NPM_TOKENsupportValidation
pnpm install --frozen-lockfilepnpm buildpnpm test