Heading is always "Apply to <Company>" on every step — comparing it
caused false stuck detection on step 2. Now:
- Progress bar selector finds <progress> elements (no explicit role)
- Stuck detection re-reads progress AFTER clicking Next to see if it changed
- Threshold raised to 3 same-progress clicks before exiting
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a required select has no answer, the unknown field now includes
the available options. Telegram notification shows them so the user
knows exactly what to reply with.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Non-EEO required selects with no answer were silently ignored, causing
infinite loops when LinkedIn blocked page advancement. Now reported as
unknown fields so the applier can exit with needs_answer status.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
page.$() and frame.$() can't find the <a> in LinkedIn's shadow DOM,
but frame.evaluateHandle with document.querySelectorAll can.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Stuck detection after clicking Next: if heading+progress unchanged 2x,
exit with 'stuck' status instead of looping forever
- Select dropdowns: treat "Select an option" as unfilled (LinkedIn's
placeholder has a truthy value that was skipping the fill logic)
- Continue button fallback: detect draft "Continue" span via apply URL
pattern when Easy Apply button not found
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- dismissModal now always waits for Discard confirmation after closing
(previously returned early after Dismiss click, leaving draft saved)
- LINKEDIN_APPLY_BUTTON_SELECTOR matches both "Easy Apply" and
"Continue applying" (shown when a previous draft exists)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
LinkedIn renders Easy Apply modal inside shadow DOM. document.querySelector()
in evaluate() cannot pierce shadow DOM, but Playwright's page.$() can.
easy_apply.mjs:
- Replaced all frame.evaluate(document.querySelector) with ElementHandle ops
- findModalButton uses modal.$$() + btn.evaluate() instead of evaluateHandle
- getModalDebugInfo uses modal.$eval and modal.$$() for all queries
- dismissModal scans buttons via page.$$() instead of evaluateHandle
- Removed findModalFrame (no longer needed)
form_filler.mjs:
- getLabel() walks up ancestor DOM to find labels (LinkedIn doesn't use label[for])
- Deduplicates repeated label text ("Phone country codePhone country code")
- isRequired() walks ancestors to find labels with * or required indicators
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
LinkedIn renders the Easy Apply modal inside a /preload/ iframe, not the
main document. page.$() searches cross-frame but evaluate() only runs in
the main frame context, causing blank headings/buttons and broken navigation.
- Added findModalFrame() to locate the frame owning the modal dialog
- All evaluate/evaluateHandle calls now use modalFrame instead of page
- findModalButton() and dismissModal() updated to accept frame parameter
- formFiller.fill() receives modalFrame so container scoping works correctly
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Logs iframe count, child element tags, and first 500 chars of modal innerHTML
on step 0 to diagnose why buttons/heading aren't being found.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Text inputs matched to cover letter now report as unknown if required,
instead of silently leaving the field empty
- Submit click now verifies modal closed before reporting success;
returns 'incomplete' with actionable log if modal stays open
- selectAutocomplete scoped to container (modal) to avoid clicking
wrong dropdowns from the underlying page
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
easy_apply.mjs:
- findModalButton() uses 3-strategy detection: aria-label exact/substring,
then exact button text match — survives LinkedIn aria-label changes
- Check order fixed: Next → Review → Submit (submit only when no forward nav)
- All queries scoped to modal + :not([disabled])
- dismissModal() with fallback chain: Dismiss → Close/X → Escape → Discard
- Uses innerText for button text (ignores hidden children)
form_filler.mjs:
- All queries scoped to container (modal when present, page otherwise)
- Radio labels use $$('label') + textContent instead of broken :has-text()
- Autocomplete uses waitForSelector instead of blind 800ms sleep
- EEO selects iterate options directly (selectOption doesn't accept regex)
- Country code check ordered before country to prevent fragile match order
constants.mjs:
- Add AUTOCOMPLETE_WAIT, AUTOCOMPLETE_TIMEOUT
- Remove unused button selectors (now handled inline by findModalButton)
ai_answer.mjs + keywords.mjs:
- Use ANTHROPIC_API_URL constant, claude-sonnet-4-6 model
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- addJobs: allows same job on multiple tracks (dedup key = track::id)
- Cross-track copies get composite id (job.id_track) to avoid batch collisions
- dedupeAfterFilter(): after collect, keeps highest-scored copy per URL, marks rest as 'duplicate'
- Called automatically at end of collect phase
- submitBatch → submitBatches: groups jobs by track, submits one batch each
- filter_state.json now stores batches[] array instead of single batch_id
- Collect waits for all batches to finish before processing
- Each track gets its own cached system prompt = better caching + cleaner scoring
- Idempotent collect: skips already-scored jobs
- Submit phase now excludes jobs with filter_batch_id set (already in a batch)
- After submitting, stamps each job with filter_batch_id = batchId
- Filtered jobs already excluded by status='filtered'
- Prevents duplicate submissions when batch errors cause state to clear without scores
- updateJobStatus was called 4,652 times causing ~4,652 file reads/writes
- Now loads queue once, applies all updates in memory, saves once
- Model was using OpenClaw alias (sonnet-4-6) not native Anthropic ID
- Only claude-3-haiku-20240307 is available on this API key; update settings.example.json
- search_runs.json: append-only history of every searcher run
(started_at, finished, added, seen, platforms, lookback_days)
- search_progress_last.json: snapshot of final progress state after
each completed run — answers 'what keywords/tracks were searched?'
- filter_runs.json: append-only history of every filter batch
(batch_id, submitted/collected timestamps, model, passed/filtered/errors)
Fixes the 'did the 90-day run complete?' ambiguity going forward
- Batch API = 50% cost savings vs synchronous calls
- Prompt caching on system prompt (profile + criteria shared across all jobs)
- One request per job with custom_id = job ID for result matching
- Two-phase state machine: submit → poll/collect (hourly cron safe)
- filter_state.json tracks pending batch ID between runs
- Model configurable via settings.filter.model (default: claude-sonnet-4-6)
- Telegram notifications on submit + collect
- Errors pass through — never block applications due to filter failure
- --stats flag for queue overview
- lib/filter.mjs: batch scoring engine (10 jobs/call, Claude Haiku)
- job_filter.mjs: standalone CLI with --dry-run and --stats flags
- Threshold configurable globally + per-search in search_config.json (filter_min_score, default 5)
- Job profiles (gtm/ae) passed as context via settings.filter.job_profiles
- Filtered jobs get status='filtered' with filter_score + filter_reason
- Filter errors pass jobs through (never block applications)
- status.mjs: added 'AI filtered' line to report
