careerclaw-js 1.0.5 → 1.1.0

package/CHANGELOG.md

@@ -1,485 +1,584 @@
-# Changelog
-
-All notable changes to careerclaw-js are documented here.
-Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
-Versioning follows [Semantic Versioning](https://semver.org/).
-
----
-
-## [Unreleased]
-
----
-
-## [1.0.5] - 2026-03-18
-
-### Fixed
-
-- Fixed hex and decimal numeric HTML entities (e.g. `/`, `/`) leaking
-  undecoded into job titles, descriptions, and outreach drafts
-  (`src/adapters/remoteok.ts`). Added generic catch-all decoding for both
-  `&#xHH;` and `&#DD;` forms before the named-entity block in `stripHtml`.
-  Triggered by the fast-xml-parser v5 upgrade (CVE-2026-26278), which stopped
-  silently pre-decoding entities inside CDATA content that v4 had handled
-  transparently.
-
----
-
-## [1.0.4] - 2026-03-08
-
-### Changed
-- Runtime check replaced — no more auto-install, just a non-blocking update nudge
-- Write-failure gate added in Step 3
-- Resume write failed row added to error table
-
----
-
-## [1.0.3] - 2026-03-08
-
-### Fixed
-- Agent no longer presents a multi-question setup form when profile is missing — redirects to resume upload only
-- Agent no longer invents career frameworks, phases, or methodologies not defined in this file
-- Agent no longer enters consultant mode during first-time setup
-
-### Changed
-- Rewrote SKILL.md from persona/behavior narrative format to numbered runbook format
-- Added hard Rules section at top with six explicit prohibitions
-- Added profile existence check (Step 1) as hard gate before any command runs
-- Setup flow now asks one question at a time — work mode first, salary second
-- Removed Agent Persona, Behavior 1, and Behavior 2 sections entirely
-- Bumped runtime self-healing check expected version to 1.0.3
-
----
-
-## [1.0.2] - 2026-03-07
-
-### Fixed
-- Behavior 2 (Strategic Gap Closing) no longer fires during First-Time Setup.
-  The agent was entering consultant mode after resume intake and ending setup
-  with open-ended targeting questions instead of proceeding to the first briefing.
-
-### Changed
-- Added explicit activation gate to Behavior 2: only applies after setup is
-  complete and only when the user asks a listed trigger phrase.
-- Step 2 of First-Time Setup now specifies a strict exit condition: collect
-  work mode and salary floor, then proceed directly to Step 3 with no
-  analysis, strategy suggestions, or targeting options.
-- Added three entries to "What Not to Do" to reinforce setup boundaries.
-
----
-
-## [1.0.1] - 2026-03-06
-
-### Changed
-- Hardened the public `SKILL.md` contract for ClawHub/OpenClaw deployment.
-- Standardized runtime behavior to a single install/execute approach:
-  - self-healing install with `npm install -g careerclaw-js@"$EXPECTED"`
-  - direct execution with `careerclaw-js ...`
-- Updated required runtime bins to match the Node/npm execution path.
-- Scoped CareerClaw behavior more clearly as an invoked skill instead of a session-wide assistant.
-- Simplified first-run onboarding around canonical `.careerclaw/resume.txt` handling.
-- Reduced first-run friction by keeping setup resume-first and minimizing follow-up questions.
-
-### Improved
-- Restored Strategic Gap Closing behavior so CareerClaw can act more like a practical career consultant when the fit is imperfect.
-- Improved recommendation tone and guidance for fatal, acceptable, and bridgeable gaps.
-- Cleaned up Pro activation flow so premium setup is only introduced when it is actually needed.
-- Restored the public Gumroad purchase link for CareerClaw Pro while keeping internal product IDs out of user-facing setup.
-
-### Removed
-- Removed internal Gumroad product ID exposure from the public skill surface.
-- Removed Python compatibility messaging from `SKILL.md` to keep the public skill contract focused on the current Node-based experience.
-
----
-
-## [1.0.0] — 2026-03-06
-
-### Added
-
-- `SKILL.md` full rewrite (Phase 12): Agent Persona section; zero-config resume
-  intake; Daily Stand-up behavior (reads `tracking.json` on session start,
-  surfaces still-live and undrafted jobs); Strategic Gap Closing (coach tone
-  for matches with gaps > 0.6 score); Sunday Night Strategy (first weekday of
-  month = HN priority, Sunday evening prompt, Monday/Friday timing context);
-  Tier-1 upsell trigger (Google, Meta, Apple, Stripe, Airbnb, Netflix);
-  Free vs Pro comparison table; `optionalEnv` frontmatter for all 12 env vars
-- `RELEASE_CHECKLIST.md` — pre-publish gate, npm publish steps, VirusTotal
-  scan instructions, OpenClaw test gate, ClawHub submission checklist
-
-### Changed
-
-- `src/llm-enhance.ts` — OpenAI dual-transport support: `gpt-4o` family uses
-  `/v1/chat/completions` with `max_completion_tokens`; newer models (gpt-5.x+)
-  use `/v1/responses` with `max_output_tokens`; transport auto-detected via
-  `defaultOpenAITransport(model)` and injectable per `ChainCandidate`;
-  `SYSTEM_PROMPT` rewritten with Bridge Sentence method as core doctrine,
-  250/300-word hard limits, 22 banned phrases by name; `buildPrompt()` updated
-  with explicit word limit and Bridge Sentence paragraph-level instruction
-- `src/config.ts` — `CAREERCLAW_DIR` empty-string guard: `??` replaced with
-  `.trim().length > 0` check so blank `CAREERCLAW_DIR=` in `.env` correctly
-  falls back to `.careerclaw/`
-- `scripts/smoke_llm.ts` — Pro key section now calls `checkLicense()` live
-  against Gumroad when both keys are set; stale "not yet implemented" message
-  removed; Polar.sh URL replaced with Gumroad URL
-- `README.md` — Polar.sh references replaced with Gumroad; CLI commands
-  corrected (no `briefing` subcommand); `--resume-pdf` removed (not
-  implemented); env vars table updated; roadmap updated through Phase 13
-
-### Notes
-
-270 tests across 16 files, all passing. `tsc --noEmit` clean.
-No new production dependencies. First ClawHub-ready release.
-OpenAI transport validated against gpt-4o-mini (chat) and gpt-5.2 (responses)
-in live smoke testing.
-
----
-
-## [0.11.0] — 2026-03-05
-
-### Added
-
-- `src/license.ts` — `checkLicense(key, options?)` → `Promise<LicenseResult>`:
-  Gumroad Pro license validation with 7-day offline cache; raw key never
-  written to disk — only `sha256(key)` cached; `{ valid: false, source: "none" }`
-  returned immediately when `CAREERCLAW_GUMROAD_PRODUCT_ID` is unset so Free
-  tier users are unaffected; `fetchFn` and `cachePath` injectable for offline tests
-- `src/tests/license.test.ts` — 12 offline tests covering the missing product ID
-  guard, fresh/stale/missing/wrong-key cache branches, all Gumroad API response
-  variants (success, refunded, chargebacked, 404, 500), never-throws invariant,
-  and hash safety (raw key must not appear in cache file)
-
-### Changed
-
-- `src/briefing.ts` — bare `proKey` presence check replaced with
-  `await checkLicense(proKey)`; `isProActive` is now driven by
-  `licenseResult.valid`; `BriefingOptions` gains `licenseFetchFn` and
-  `licenseCachePath` for test injection
-- `src/config.ts` — added `GUMROAD_PRODUCT_ID`, `GUMROAD_API_BASE`,
-  `LICENSE_CACHE_TTL_MS` (7 days in ms); `POLAR_PRODUCT_SLUG` and
-  `POLAR_API_BASE` retained and marked `@deprecated` with
-  `TODO: Phase 11-Polar` migration comment
-- `.env.example` — Pro purchase URL updated to Gumroad;
-  `CAREERCLAW_GUMROAD_PRODUCT_ID=` added with dashboard instructions
-
-### Security
-
-- Raw license key is never written to disk; only `sha256(key)` is persisted
-- Gumroad validation uses `increment_uses_count=false` — validation calls
-  do not consume customer usage quota
-
-### Notes
-
-270 tests across 16 files, all passing. `tsc --noEmit` clean.
-No new production dependencies. Polar migration path preserved via
-`@deprecated` constants and `TODO: Phase 11-Polar` comment in `config.ts`.
-
----
-
-## [0.10.0] — 2026-03-05
-
-### Added
-
-- `src/llm-enhance.ts` — `enhanceDraft(job, profile, resumeIntel, draft, gapKeywords, options)`:
-  Pro tier LLM-powered outreach draft generation; parses `LLM_CHAIN` into a left-to-right failover
-  list; per-candidate retry loop up to `LLM_MAX_RETRIES`; circuit breaker opens after
-  `LLM_CIRCUIT_BREAKER_FAILS` consecutive failures; `_chainOverride` for offline testing
-- `src/tests/llm-enhance.test.ts` — 12 offline tests covering Anthropic path, OpenAI path,
-  HTTP error fallback, unparseable response fallback, never-throws invariant, circuit breaker
-  call-count bounds, privacy (raw resume text must not appear in outbound request payload),
-  and empty-chain no-op
-
-### Changed
-
-- `src/briefing.ts` — `BriefingOptions` gains `resumeIntel?: ResumeIntelligence`,
-  `proKey?: string`, `enhanceFetchFn?`; draft stage converted to `async`; calls `enhanceDraft()`
-  per match when `proKey` is set — falls back to deterministic baseline on any LLM failure
-- `src/cli.ts` — calls `buildResumeIntelligence()` after resume load; passes `resumeIntel` and
-  `PRO_KEY` to `runBriefing()`; Pro enhancement path is now live from the CLI
-
-### Security
-
-- LLM prompt contains only extracted keyword signals (`impact_signals` ≤ 12 tokens,
-  `gap_keywords` ≤ 6 tokens) — raw resume text and `resume_summary` are never sent to
-  any external API
-
-### Notes
-
-258 tests across 15 files, all passing. `tsc --noEmit` clean. No new production dependencies.
-License validation (`CAREERCLAW_PRO_KEY` checked against Gumroad) is deferred to Phase 11 (v0.11.0) —
-`proKey` is trusted by presence only in this version.
-
----
-
-## [0.8.1] — 2026-03-04
-
-### Fixed
-
-- **Dentist problem**: replaced additive composite score with a multiplicative model (`total = sqrt(keyword) × qualityBase`);
-  zero keyword overlap now always produces a score of 0.0 regardless of metadata alignment — irrelevant jobs can no longer float to the
-  top on neutral dimension scores
-- Signal gate added to `rankJobs()`: jobs with keyword score below `minKeywordScore` (default: 0.01) are hard-filtered before ranking
-- `matched_keywords` and `gap_keywords` were hardcoded empty in `engine.ts`; now wired from `compositeScore()` output
-- HTML entity `&#x2F;` (and all `&#x[hex];` sequences) now decoded in `stripHtml()` — fixes `ML&#x2F;AI` appearing in HN job titles and
-  gap tokens
-- `stripHtml()` now applied to parsed HN job title and company name (previously body text only)
-- Contraction tokens (`i'm`, `i've`, `don't`, etc.) added to `STOPWORDS` — eliminates noise in gap keyword output
-
-### Changed
-
-- `MatchBreakdown` field names renamed: `keyword_score → keyword`, `experience_score → experience`, `salary_score → salary`,
-  `work_mode_score → work_mode`
-- `compositeScore()` return type extended: now includes `matched` and `gaps` string arrays alongside `total` and `breakdown`
-- Default LLM model updated: `claude-sonnet-4-20250514` → `claude-haiku-4-5-20251001`
-- `scripts/smoke_briefing.ts` refactored to multi-profile mode; select profile via `PROFILE=0|1|2` env var
-
-### Added
-
-- `.env.example` — full credential reference (OpenClaw gateway, agent LLM keys, Pro license, draft enhancement keys, failover chain config)
-- `SECURITY.md` — local-first security architecture, credential handling policy, external network call inventory, LLM data disclosure
-- `src/config.ts`: `LLM_ANTHROPIC_KEY`, `LLM_OPENAI_KEY`, `LLM_CHAIN`, `LLM_MAX_RETRIES`, `LLM_CIRCUIT_BREAKER_FAILS`
-- `scripts/smoke_llm.ts` — LLM connectivity and Pro key smoke test; `npm run smoke:llm`
-
-### Notes
-
-234 tests across 13 files, all passing. No new production dependencies.
-`MatchBreakdown` rename is a breaking internal change — no public API consumers exist prior to the CLI (Phase 9).
-
----
-
-## [0.8.0] — 2026-03-04
-
-### Added
-
-- `src/briefing.ts` — `runBriefing(profile, options)`: end-to-end pipeline orchestrator; four timed stages (fetch, rank, draft,
-  persist); skill-first design — profile is a parameter, never loaded from disk; `fetchFn` and `repo` are injectable for testing; dry-run
-  suppresses all writes while keeping counts accurate; catastrophic fetch failure returns empty result rather than throwing;
-  `run_id` generated via `crypto.randomUUID()` (Node built-in); `version` read from package.json at runtime via `createRequire`
-- `BriefingResult` interface added to `src/models.ts` — stable JSON output schema for OpenClaw/ClawHub agent consumption: `run`,
-  `matches`, `drafts`, `tracking`, `dry_run`
-- `src/tests/briefing.test.ts` — 17 integration tests; all offline via stubbed `fetchFn` and `tmpdir`-backed `TrackingRepository`
-
-### Fixed
-
-- `draftOutreach()` and `upsertEntries()` were being passed `ScoredJob` where `NormalizedJob` was required; `ScoredJob` wraps `NormalizedJob`
-  as `.job` and does not extend it — both callsites corrected to unwrap `.job`; corresponding test assertion corrected from
-  `scored.job_id` to `scored.job.job_id`
-
-### Notes
-
-243 tests across 13 files, all passing. No new dependencies. The pipeline is now end-to-end complete. Every module from Phase 1 through
-Phase 7 is wired into a single callable function. Phase 9 (CLI entry point) will expose `runBriefing()` as an executable with `--dry-run`,
-`--top-k`, and `--json` flags, matching the Python careerclaw CLI surface.
-
----
-
-## [0.7.0] — 2026-03-04
-
-### Added
-
-- `src/tracking.ts` — `TrackingRepository` class managing two runtime files under `.careerclaw/`:
-  - `tracking.json` — keyed object of `TrackingEntry` records; new jobs saved with status `"saved"`; re-encountered jobs have
-    `last_seen_at` and `updated_at` refreshed, all other fields (including user-set status) preserved; one disk write per
-    `upsertEntries()` call regardless of batch size
-  - `runs.jsonl` — append-only newline-delimited JSON log; one `BriefingRun` per line via `appendRun()`
-  - Constructor accepts `trackingPath`, `runsPath`, and `dryRun` overrides; defaults from `config.ts`
-  - `load()` returns empty store on a missing or corrupt file — no crash
-  - `ensureDir()` creates parent directory recursively on first writing
-  - `upsertEntries()` returns accurate `{ created, already_present }` counts even in dry-run mode
-  - `ScoredJob?` parameter on `upsertEntries()` reserved for Phase 8 score snapshot attachment
-- `src/tests/tracking.test.ts` — 23 unit tests; per-test isolated `tmpdir` directories; covers load, upsert (new + re-encounter),
-  disk writes, JSONL append, and dry-run suppression
-
-### Notes
-
-226 tests across 12 files, all passing. No new dependencies. The `TrackingRepository` is the last infrastructure module before the
-Phase 8 briefing orchestrator wires everything into the full pipeline.
-Dry-run mode is a first-class concern throughout: all writing paths are suppressed, all read and count paths remain fully functional, matching
-the Python careerclaw `--dry-run` behaviour.
-
----
-
-## [0.6.0] — 2026-03-04
-
-### Added
-
-- `src/drafting.ts` — `draftOutreach(job, profile, matchedKeywords)`:
-  deterministic outreach email generator; subject line follows `Interest in {title} at {company}` format; body inserts experience
-  clause (years or "extensive experience" fallback), up to 3 matched keywords formatted as natural language, and 3 fixed
-  reliability/collaboration/instrumentation bullet highlights; word count 161–168 words depending on a keyword path, inside 150–250 word
-  spec; `llm_enhanced: false` always; `formatList()` helper for natural-language list formatting
-- `src/tests/drafting.test.ts` — 20 unit tests
-
-### Fixed
-
-- Deterministic template body was 127 words on the first pass — below the 150-word MVP spec floor; fixed by expanding opening and closing
-  paragraphs; both keyword and fallback paths re-verified at 161 and 168 words respectively
-
-### Notes
-
-203 tests across 11 files, all passing. No new dependencies. `llm_enhanced` is always false in this phase — LLM enhancement
-(Phase 7+) will set this flag to true when the Pro key is configured and the call succeeds. The deterministic template remains the permanent
-fallback for the Free tier and for LLM failure scenarios.
-
----
-
-## [0.5.0] — 2026-03-04
-
-### Added
-
-- `src/requirements.ts` — `extractJobRequirements(job)`: tokenizes job title and description into a deduplicated keyword and phrase corpus for
-  use as the job-side input to gap analysis
-- `src/resume-intel.ts` — `buildResumeIntelligence(params)`:
-  section-aware keyword/phrase extraction across skills (weight 1.0), summary + target_roles (weight 0.8), and optional resume_text (weight
-  0.6); per-keyword weight is the max across sections; `impact_signals` are keywords with weight >= 0.8; `source` flag indicates which inputs
-  contributed; PR-E fix (skills injection) baked in from day one
-- `src/gap.ts` — `gapAnalysis(intel, job)`: weighted `fit_score` (sum of matched keyword_weights / job keyword count), `fit_score_unweighted`
-  (Jaccard), `signals` (resume ∩ job), `gaps` (job − resume), and top-5 `summary` for display
-- `JobRequirements`, `ResumeIntelligence`, `GapAnalysisResult` interfaces added to `src/models.ts`; `ResumeIntelligence` schema is
-  JSON-compatible with Python careerclaw output
-- `src/tests/resume-intel.test.ts` — 19 unit tests
-- `src/tests/gap.test.ts` — 16 unit tests
-
-### Fixed
-
-- Added `"am"` to `STOPWORDS` in `src/core/text-processing.ts` — missed from an initial set alongside `"is"`, `"are"`, `"was"`, `"were"`, `"be"`;
-  caught by resume-intel stopword filter test
-
-### Notes
-
-183 tests across 10 files, all passing. No new dependencies. The `fit_score` weighted formula is identical to the Python careerclaw
-implementation: skills listed in UserProfile. Skills receive weight 1.0 and will never appear as gaps. The practical fit_score ceiling against
-real job postings is ~50% due to company names and location tokens in the denominator.
-
-### Future Work
-
-- CorpusCache: Entropy-based token filtering (IDF) to suppress tokens that appear in >80% of fetched jobs. Gated behind corpus_size >= 50.
-  Planned for a future release after job tracking accumulates sufficient data.
-
----
-
-## [0.4.0] — 2026-03-04
-
-### Added
-
-- `src/matching/scoring.ts` — four pure scoring functions:
-  `scoreKeyword()` (Jaccard token overlap, returns matched and gap keyword lists), `scoreExperience()` (clamped linear user/job years ratio),
-  `scoreSalary()` (proportional against a user minimum),
-  `scoreWorkMode()` (exact=1.0, hybrid=0.5 partial, mismatch=0.0);
-  `compositeScore()` with `WEIGHTS` (keyword=0.50, experience=0.20, salary=0.15, work_mode=0.15)
-- `src/matching/engine.ts` — `rankJobs(jobs, profile, topK)` scores all jobs, sorts descending by composite score, returns top-K `ScoredJob[]`
-  with full breakdown and keyword lists; scores rounded to 4 d.p.
-- `src/matching/index.ts` — barrel export for matching public API
-- `src/tests/matching.scoring.test.ts` — 36 unit tests
-- `src/tests/matching.engine.test.ts` — 10 end-to-end tests using real model types
-
-### Notes
-
-148 tests across 8 files, all passing. No new dependencies. Neutral score (0.5) is used for all null data cases, so missing job fields
-neither reward nor penalize the composite — consistent with Python careerclaw behavior. Gap keywords from `scoreKeyword()` feed directly
-into Phase 5 gap analysis.
-
----
-
-## [0.3.0] — 2026-03-04
-
-### Added
-
-- `src/sources.ts` — source aggregation layer: `fetchAllJobs()` runs both
-  adapters concurrently with per-source error isolation; `deduplicate()`
-  removes duplicate `job_id` entries (first-seen wins); returns `FetchResult`
-  with job list, per-source counts, and error map for run instrumentation
-- `src/core/text-processing.ts` — shared text processing library:
-  `STOPWORDS` (English function words and full PR-E recruitment boilerplate set),
-  `SECTION_WEIGHTS` (skills=1.0, summary=0.8, experience=0.6, education=0.4),
-  `tokenize()`, `tokenizeUnique()`, `extractPhrases()`,
-  `extractPhrasesFromText()`, `tokenOverlap()`, `matchedTokens()`,
-  `gapTokens()`
-- `src/tests/text-processing.test.ts` — 34 unit tests
-- `src/tests/sources.test.ts` — 10 unit tests (ESM-safe adapter stubs via
-  `vi.doMock()` + `vi.resetModules()`; no network)
-
-### Notes
-
-102 tests across 6 files, all passing. No new production dependencies.
-`SECTION_WEIGHTS` is defined here and will be consumed by resume intelligence
-in Phase 5. The `FetchResult.errors` map feeds into `BriefingRun.sources`
-instrumentation in Phase 8.
-
----
-
-## [0.2.0] — 2026-03-03
-
-### Added
-
-- `src/adapters/remoteok.ts` — RemoteOK RSS adapter; parses RSS XML into
-  `NormalizedJob[]`; `parseRss()` exported separately from `fetchRemoteOkJobs()`
-  so contract tests call pure parsing functions without network mocking
-- `src/adapters/hackernews.ts` — HN Firebase adapter; fetches "Who is Hiring?"
-  thread comments in parallel; `parseComment()` exported for offline testing;
-  handles deleted/dead items gracefully
-- `src/adapters/index.ts` — barrel export for all adapter public API
-- `src/tests/fixtures/remoteok.xml` — RSS fixture covering full fields, no-salary,
-  and k-suffix salary variants
-- `src/tests/fixtures/hn-thread.json` — HN thread fixture with `kids` array
-- `src/tests/fixtures/hn-comment-job.json` — HN job comment fixture (pipe-separated
-  header, HTML body, salary, experience years)
-- `src/tests/fixtures/hn-comment-deleted.json` — deleted comment fixture (adapter
-  must skip)
-- `src/tests/adapters.remoteok.test.ts` — 25 offline contract tests (title/company
-  splitting, salary parsing, work-mode inference, HTML stripping, `stableId`)
-- `src/tests/adapters.hackernews.test.ts` — 18 offline contract tests (header
-  parsing, timestamp conversion, HTML decoding, skip logic for deleted items)
-- `scripts/smoke_sources.ts` — live smoke test hitting real RemoteOK RSS and HN
-  Firebase APIs; run manually before releases with `npm run smoke`
-- `fast-xml-parser` added as a production dependency (RSS parsing)
-- `tsx` added as a dev dependency (runs a smoke script without a compiler step)
-
-### Changed
-
-- `stripHtml()` fixed: opening `<p>` tags now convert to `\n` (was `""`) so HN
-  comment header and body lines split correctly after HTML stripping
-- `README.md` — roadmap updated; Phase 2 marked complete; note updated to
-  reference v0.2.0
-- **Payment processor:** Pro license switched from Gumroad to **Polar.sh**
-  (`https://polar.sh/orestes-garcia-martinez/careerclaw-pro`); `CAREERCLAW_PRO_KEY`
-  env var name and SHA-256 cache behavior unchanged
-
-### Notes
-
-58 tests across 4 test files, all passing. No network calls in CI — all adapter
-tests use offline fixtures. Run `npm run smoke` manually before each release to
-validate live sources.
-
-
-## [0.1.0] — 2026-03-03
-
-### Added
-
-- Initial repository scaffold: `package.json`, `tsconfig.json`, `vitest.config.ts`
-- `src/models.ts` — canonical data schemas (`NormalizedJob`, `UserProfile`,
-  `TrackingEntry`, `BriefingRun`, `ScoredJob`, `OutreachDraft`); identical
-  JSON serialization format to Python careerclaw for cross-implementation
-  file compatibility
-- `src/config.ts` — centralised environment and source configuration
-  (runtime paths, HTTP defaults, RemoteOK RSS URL, HN thread ID, LLM and
-  license env vars)
-- `SKILL.md` — OpenClaw skill definition with Node-native self-healing
-  install check (`npm install -g careerclaw-js`)
-- `CHANGELOG.md`
-- Unit tests for `models.ts` and `config.ts` (Vitest)
-
-### Notes
-
-This release establishes the Phase 1 foundation types. No adapters,
-matching, or CLI are included yet — those follow in Phases 2–8 per the
-Node Migration Decision (ADR, March 2026).
-
-[Unreleased]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.8.0...HEAD
-[0.8.1]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.8.0...v0.8.1
-[0.8.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.7.0...v0.8.0
-[0.7.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.6.0...v0.7.0
-[0.6.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.5.0...v0.6.0
-[0.5.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.4.0...v0.5.0
-[0.4.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.3.0...v0.4.0
-[0.3.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.2.0...v0.3.0
-[0.2.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.1.0...v0.2.0
+# Changelog
+
+All notable changes to careerclaw-js are documented here.
+Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+Versioning follows [Semantic Versioning](https://semver.org/).
+
+---
+
+## [Unreleased]
+
+---
+
+## [1.1.0] - 2026-03-26
+
+### Added
+
+- `src/execution-context.ts` — dual execution-context architecture for CareerClaw
+  runtimes; `ClawOsExecutionContext` (trusted platform path, `verified: true`,
+  feature list passed from upstream entitlement) and `StandaloneExecutionContext`
+  (local CLI / direct package use with Gumroad license validation);
+  `CAREERCLAW_FEATURES` constant registry with four feature flags
+  (`LLM_OUTREACH_DRAFT`, `TAILORED_COVER_LETTER`, `RESUME_GAP_ANALYSIS`,
+  `TOPK_EXTENDED`); `createClawOsExecutionContext()` and
+  `createStandaloneExecutionContext()` factory functions;
+  `hasCareerClawFeature()` type-safe feature check helper;
+  `CareerClawFeatureKey` union type derived from `CAREERCLAW_FEATURES`
+- `src/runtime.ts` — programmatic runtime wrappers providing the clean
+  import boundary for both ClawOS and standalone consumers;
+  `runCareerClawStandalone(input, options)` delegates to `runBriefing()` with
+  standalone license flow; `runCareerClawWithContext(input, context, options)`
+  delegates to `runBriefingWithContext()` with trusted platform context;
+  `resolveResumeIntelligence()` centralises resume-intel construction from
+  profile + optional resume text, eliminating duplicate logic across callers;
+  `CareerClawRunInput`, `CareerClawRunSupportOptions`, and
+  `CareerClawStandaloneRunOptions` interfaces for typed consumer APIs
+- `src/index.ts` — public API barrel export; re-exports `models`, `briefing`,
+  `runtime`, `resume-intel`, `execution-context`, `license`, `tracking`, and
+  `sources` modules; enables `import { ... } from "careerclaw-js"` for all
+  public types and functions
+- `src/tests/runtime.test.ts` — 4 integration tests covering the trust
+  boundary security matrix: standalone with Pro key (license validated via
+  Gumroad → LLM drafts active), standalone without Pro key (no license check →
+  template drafts only), ClawOS with `LLM_OUTREACH_DRAFT` feature (no Gumroad
+  call → LLM drafts active), ClawOS without feature (no Gumroad call → template
+  drafts only); mocks `checkLicense` and `enhanceDraft` to verify call isolation
+
+### Changed
+
+- `src/briefing.ts` — refactored from single `runBriefing()` entry point to
+  dual-mode architecture via internal strategy pattern; added
+  `runBriefingWithContext(profile, context, options)` as the trusted ClawOS
+  entry point; extracted `runBriefingInternal()` as the shared pipeline with
+  `InternalExecutionMode` discriminated union (`"standalone"` | `"clawos"`);
+  extracted `resolvePremiumDraftAccess(executionMode, resumeIntel)` to resolve
+  Pro activation per mode — standalone validates via `checkLicense()`, ClawOS
+  checks `tier === "pro"` + `hasCareerClawFeature(LLM_OUTREACH_DRAFT)`;
+  fixed unsafe `resumeIntel!` non-null assertion to safe `isProActive && resumeIntel`
+  guard; added `ContextBriefingOptions` interface (excludes `proKey`,
+  `licenseFetchFn`, `licenseCachePath` — not applicable in trusted mode);
+  removed inline JSDoc comments from `BriefingOptions` fields (types are
+  self-documenting)
+- `src/cli.ts` — added `--version` / `-v` flag using `createRequire` to read
+  version from `package.json`; version check exits before any profile or config
+  loading; updated `printHelp()` to document the new flag; replaced direct
+  `runBriefing()` call with `runCareerClawStandalone()` from `runtime.ts`,
+  delegating resume-intel construction and option wiring to the runtime layer
+- `package.json` — description updated to reflect dual-runtime identity
+  (ClawOS + standalone); `"clawos"` keyword added; `exports` map added with
+  three subpath entries: `"."` (main library, `dist/index.js`), `"./cli"`
+  (CLI entry, `dist/cli.js`), `"./package.json"` (self-reference for
+  `createRequire` version reads)
+- `README.md` — full rewrite aligned with dual execution-context architecture;
+  added [ClawOS](https://clawoshq.com/) links throughout; added
+  [ClawHub](https://clawhub.ai/) and [OpenClaw](https://openclaw.org/)
+  references for standalone skill distribution; split Pro pricing into two
+  channels (ClawOS $9/month recommended, Standalone $39 lifetime via Gumroad);
+  added programmatic integration examples for both `runCareerClawWithContext()`
+  (with `createClawOsExecutionContext()`) and `runCareerClawStandalone()`;
+  added `--version` flag to CLI options; updated project structure tree with
+  `execution-context.ts`, `index.ts`, and `runtime.ts`
+
+### Security
+
+- ClawOS execution path never touches Gumroad — entitlement is resolved
+  upstream by the platform; the `ClawOsExecutionContext.verified: true`
+  literal type enforces that only pre-verified contexts are accepted
+- Standalone execution path always validates through `checkLicense()` when a
+  `proKey` is provided — no bypass flag exists on the public CLI
+- `resumeIntel!` non-null assertion removed from the draft enhancement call;
+  replaced with an explicit `isProActive && resumeIntel` guard to prevent
+  runtime errors if `resumeIntel` is undefined
+
+### Notes
+
+274 tests across 17 files, all passing. `tsc --noEmit` clean.
+No new production dependencies. The `exports` map in `package.json` makes
+the package a proper ESM citizen with subpath resolution for ClawOS
+direct-import consumption.
+
+---
+
+## [1.0.6] - 2026-03-19
+
+### Fixed
+
+- Triggered by the fast-xml-parser 5.5.7, upgrade (CVE-2026-33349), 
+  which Entity Expansion Limits Bypassed When Set to Zero Due to JavaScript 
+  Falsy Evaluation in fast-xml-parser
+
+---
+
+## [1.0.5] - 2026-03-18
+
+### Fixed
+
+- Fixed hex and decimal numeric HTML entities (e.g. `&#x2F;`, `&#47;`) leaking
+  undecoded into job titles, descriptions, and outreach drafts
+  (`src/adapters/remoteok.ts`). Added generic catch-all decoding for both
+  `&#xHH;` and `&#DD;` forms before the named-entity block in `stripHtml`.
+  Triggered by the fast-xml-parser v5 upgrade (CVE-2026-26278), which stopped
+  silently pre-decoding entities inside CDATA content that v4 had handled
+  transparently.
+
+---
+
+## [1.0.4] - 2026-03-08
+
+### Changed
+- Runtime check replaced — no more auto-install, just a non-blocking update nudge
+- Write-failure gate added in Step 3
+- Resume write failed row added to error table
+
+---
+
+## [1.0.3] - 2026-03-08
+
+### Fixed
+- Agent no longer presents a multi-question setup form when profile is missing — redirects to resume upload only
+- Agent no longer invents career frameworks, phases, or methodologies not defined in this file
+- Agent no longer enters consultant mode during first-time setup
+
+### Changed
+- Rewrote SKILL.md from persona/behavior narrative format to numbered runbook format
+- Added hard Rules section at top with six explicit prohibitions
+- Added profile existence check (Step 1) as hard gate before any command runs
+- Setup flow now asks one question at a time — work mode first, salary second
+- Removed Agent Persona, Behavior 1, and Behavior 2 sections entirely
+- Bumped runtime self-healing check expected version to 1.0.3
+
+---
+
+## [1.0.2] - 2026-03-07
+
+### Fixed
+- Behavior 2 (Strategic Gap Closing) no longer fires during First-Time Setup.
+  The agent was entering consultant mode after resume intake and ending setup
+  with open-ended targeting questions instead of proceeding to the first briefing.
+
+### Changed
+- Added explicit activation gate to Behavior 2: only applies after setup is
+  complete and only when the user asks a listed trigger phrase.
+- Step 2 of First-Time Setup now specifies a strict exit condition: collect
+  work mode and salary floor, then proceed directly to Step 3 with no
+  analysis, strategy suggestions, or targeting options.
+- Added three entries to "What Not to Do" to reinforce setup boundaries.
+
+---
+
+## [1.0.1] - 2026-03-06
+
+### Changed
+- Hardened the public `SKILL.md` contract for ClawHub/OpenClaw deployment.
+- Standardized runtime behavior to a single install/execute approach:
+  - self-healing install with `npm install -g careerclaw-js@"$EXPECTED"`
+  - direct execution with `careerclaw-js ...`
+- Updated required runtime bins to match the Node/npm execution path.
+- Scoped CareerClaw behavior more clearly as an invoked skill instead of a session-wide assistant.
+- Simplified first-run onboarding around canonical `.careerclaw/resume.txt` handling.
+- Reduced first-run friction by keeping setup resume-first and minimizing follow-up questions.
+
+### Improved
+- Restored Strategic Gap Closing behavior so CareerClaw can act more like a practical career consultant when the fit is imperfect.
+- Improved recommendation tone and guidance for fatal, acceptable, and bridgeable gaps.
+- Cleaned up Pro activation flow so premium setup is only introduced when it is actually needed.
+- Restored the public Gumroad purchase link for CareerClaw Pro while keeping internal product IDs out of user-facing setup.
+
+### Removed
+- Removed internal Gumroad product ID exposure from the public skill surface.
+- Removed Python compatibility messaging from `SKILL.md` to keep the public skill contract focused on the current Node-based experience.
+
+---
+
+## [1.0.0] — 2026-03-06
+
+### Added
+
+- `SKILL.md` full rewrite (Phase 12): Agent Persona section; zero-config resume
+  intake; Daily Stand-up behavior (reads `tracking.json` on session start,
+  surfaces still-live and undrafted jobs); Strategic Gap Closing (coach tone
+  for matches with gaps > 0.6 score); Sunday Night Strategy (first weekday of
+  month = HN priority, Sunday evening prompt, Monday/Friday timing context);
+  Tier-1 upsell trigger (Google, Meta, Apple, Stripe, Airbnb, Netflix);
+  Free vs Pro comparison table; `optionalEnv` frontmatter for all 12 env vars
+- `RELEASE_CHECKLIST.md` — pre-publish gate, npm publish steps, VirusTotal
+  scan instructions, OpenClaw test gate, ClawHub submission checklist
+
+### Changed
+
+- `src/llm-enhance.ts` — OpenAI dual-transport support: `gpt-4o` family uses
+  `/v1/chat/completions` with `max_completion_tokens`; newer models (gpt-5.x+)
+  use `/v1/responses` with `max_output_tokens`; transport auto-detected via
+  `defaultOpenAITransport(model)` and injectable per `ChainCandidate`;
+  `SYSTEM_PROMPT` rewritten with Bridge Sentence method as core doctrine,
+  250/300-word hard limits, 22 banned phrases by name; `buildPrompt()` updated
+  with explicit word limit and Bridge Sentence paragraph-level instruction
+- `src/config.ts` — `CAREERCLAW_DIR` empty-string guard: `??` replaced with
+  `.trim().length > 0` check so blank `CAREERCLAW_DIR=` in `.env` correctly
+  falls back to `.careerclaw/`
+- `scripts/smoke_llm.ts` — Pro key section now calls `checkLicense()` live
+  against Gumroad when both keys are set; stale "not yet implemented" message
+  removed; Polar.sh URL replaced with Gumroad URL
+- `README.md` — Polar.sh references replaced with Gumroad; CLI commands
+  corrected (no `briefing` subcommand); `--resume-pdf` removed (not
+  implemented); env vars table updated; roadmap updated through Phase 13
+
+### Notes
+
+270 tests across 16 files, all passing. `tsc --noEmit` clean.
+No new production dependencies. First ClawHub-ready release.
+OpenAI transport validated against gpt-4o-mini (chat) and gpt-5.2 (responses)
+in live smoke testing.
+
+---
+
+## [0.11.0] — 2026-03-05
+
+### Added
+
+- `src/license.ts` — `checkLicense(key, options?)` → `Promise<LicenseResult>`:
+  Gumroad Pro license validation with 7-day offline cache; raw key never
+  written to disk — only `sha256(key)` cached; `{ valid: false, source: "none" }`
+  returned immediately when `CAREERCLAW_GUMROAD_PRODUCT_ID` is unset so Free
+  tier users are unaffected; `fetchFn` and `cachePath` injectable for offline tests
+- `src/tests/license.test.ts` — 12 offline tests covering the missing product ID
+  guard, fresh/stale/missing/wrong-key cache branches, all Gumroad API response
+  variants (success, refunded, chargebacked, 404, 500), never-throws invariant,
+  and hash safety (raw key must not appear in cache file)
+
+### Changed
+
+- `src/briefing.ts` — bare `proKey` presence check replaced with
+  `await checkLicense(proKey)`; `isProActive` is now driven by
+  `licenseResult.valid`; `BriefingOptions` gains `licenseFetchFn` and
+  `licenseCachePath` for test injection
+- `src/config.ts` — added `GUMROAD_PRODUCT_ID`, `GUMROAD_API_BASE`,
+  `LICENSE_CACHE_TTL_MS` (7 days in ms); `POLAR_PRODUCT_SLUG` and
+  `POLAR_API_BASE` retained and marked `@deprecated` with
+  `TODO: Phase 11-Polar` migration comment
+- `.env.example` — Pro purchase URL updated to Gumroad;
+  `CAREERCLAW_GUMROAD_PRODUCT_ID=` added with dashboard instructions
+
+### Security
+
+- Raw license key is never written to disk; only `sha256(key)` is persisted
+- Gumroad validation uses `increment_uses_count=false` — validation calls
+  do not consume customer usage quota
+
+### Notes
+
+270 tests across 16 files, all passing. `tsc --noEmit` clean.
+No new production dependencies. Polar migration path preserved via
+`@deprecated` constants and `TODO: Phase 11-Polar` comment in `config.ts`.
+
+---
+
+## [0.10.0] — 2026-03-05
+
+### Added
+
+- `src/llm-enhance.ts` — `enhanceDraft(job, profile, resumeIntel, draft, gapKeywords, options)`:
+  Pro tier LLM-powered outreach draft generation; parses `LLM_CHAIN` into a left-to-right failover
+  list; per-candidate retry loop up to `LLM_MAX_RETRIES`; circuit breaker opens after
+  `LLM_CIRCUIT_BREAKER_FAILS` consecutive failures; `_chainOverride` for offline testing
+- `src/tests/llm-enhance.test.ts` — 12 offline tests covering Anthropic path, OpenAI path,
+  HTTP error fallback, unparseable response fallback, never-throws invariant, circuit breaker
+  call-count bounds, privacy (raw resume text must not appear in outbound request payload),
+  and empty-chain no-op
+
+### Changed
+
+- `src/briefing.ts` — `BriefingOptions` gains `resumeIntel?: ResumeIntelligence`,
+  `proKey?: string`, `enhanceFetchFn?`; draft stage converted to `async`; calls `enhanceDraft()`
+  per match when `proKey` is set — falls back to deterministic baseline on any LLM failure
+- `src/cli.ts` — calls `buildResumeIntelligence()` after resume load; passes `resumeIntel` and
+  `PRO_KEY` to `runBriefing()`; Pro enhancement path is now live from the CLI
+
+### Security
+
+- LLM prompt contains only extracted keyword signals (`impact_signals` ≤ 12 tokens,
+  `gap_keywords` ≤ 6 tokens) — raw resume text and `resume_summary` are never sent to
+  any external API
+
+### Notes
+
+258 tests across 15 files, all passing. `tsc --noEmit` clean. No new production dependencies.
+License validation (`CAREERCLAW_PRO_KEY` checked against Gumroad) is deferred to Phase 11 (v0.11.0) —
+`proKey` is trusted by presence only in this version.
+
+---
+
+## [0.8.1] — 2026-03-04
+
+### Fixed
+
+- **Dentist problem**: replaced additive composite score with a multiplicative model (`total = sqrt(keyword) × qualityBase`);
+  zero keyword overlap now always produces a score of 0.0 regardless of metadata alignment — irrelevant jobs can no longer float to the
+  top on neutral dimension scores
+- Signal gate added to `rankJobs()`: jobs with keyword score below `minKeywordScore` (default: 0.01) are hard-filtered before ranking
+- `matched_keywords` and `gap_keywords` were hardcoded empty in `engine.ts`; now wired from `compositeScore()` output
+- HTML entity `&#x2F;` (and all `&#x[hex];` sequences) now decoded in `stripHtml()` — fixes `ML&#x2F;AI` appearing in HN job titles and
+  gap tokens
+- `stripHtml()` now applied to parsed HN job title and company name (previously body text only)
+- Contraction tokens (`i'm`, `i've`, `don't`, etc.) added to `STOPWORDS` — eliminates noise in gap keyword output
+
+### Changed
+
+- `MatchBreakdown` field names renamed: `keyword_score → keyword`, `experience_score → experience`, `salary_score → salary`,
+  `work_mode_score → work_mode`
+- `compositeScore()` return type extended: now includes `matched` and `gaps` string arrays alongside `total` and `breakdown`
+- Default LLM model updated: `claude-sonnet-4-20250514` → `claude-haiku-4-5-20251001`
+- `scripts/smoke_briefing.ts` refactored to multi-profile mode; select profile via `PROFILE=0|1|2` env var
+
+### Added
+
+- `.env.example` — full credential reference (OpenClaw gateway, agent LLM keys, Pro license, draft enhancement keys, failover chain config)
+- `SECURITY.md` — local-first security architecture, credential handling policy, external network call inventory, LLM data disclosure
+- `src/config.ts`: `LLM_ANTHROPIC_KEY`, `LLM_OPENAI_KEY`, `LLM_CHAIN`, `LLM_MAX_RETRIES`, `LLM_CIRCUIT_BREAKER_FAILS`
+- `scripts/smoke_llm.ts` — LLM connectivity and Pro key smoke test; `npm run smoke:llm`
+
+### Notes
+
+234 tests across 13 files, all passing. No new production dependencies.
+`MatchBreakdown` rename is a breaking internal change — no public API consumers exist prior to the CLI (Phase 9).
+
+---
+
+## [0.8.0] — 2026-03-04
+
+### Added
+
+- `src/briefing.ts` — `runBriefing(profile, options)`: end-to-end pipeline orchestrator; four timed stages (fetch, rank, draft,
+  persist); skill-first design — profile is a parameter, never loaded from disk; `fetchFn` and `repo` are injectable for testing; dry-run
+  suppresses all writes while keeping counts accurate; catastrophic fetch failure returns empty result rather than throwing;
+  `run_id` generated via `crypto.randomUUID()` (Node built-in); `version` read from package.json at runtime via `createRequire`
+- `BriefingResult` interface added to `src/models.ts` — stable JSON output schema for OpenClaw/ClawHub agent consumption: `run`,
+  `matches`, `drafts`, `tracking`, `dry_run`
+- `src/tests/briefing.test.ts` — 17 integration tests; all offline via stubbed `fetchFn` and `tmpdir`-backed `TrackingRepository`
+
+### Fixed
+
+- `draftOutreach()` and `upsertEntries()` were being passed `ScoredJob` where `NormalizedJob` was required; `ScoredJob` wraps `NormalizedJob`
+  as `.job` and does not extend it — both callsites corrected to unwrap `.job`; corresponding test assertion corrected from
+  `scored.job_id` to `scored.job.job_id`
+
+### Notes
+
+243 tests across 13 files, all passing. No new dependencies. The pipeline is now end-to-end complete. Every module from Phase 1 through
+Phase 7 is wired into a single callable function. Phase 9 (CLI entry point) will expose `runBriefing()` as an executable with `--dry-run`,
+`--top-k`, and `--json` flags, matching the Python careerclaw CLI surface.
+
+---
+
+## [0.7.0] — 2026-03-04
+
+### Added
+
+- `src/tracking.ts` — `TrackingRepository` class managing two runtime files under `.careerclaw/`:
+  - `tracking.json` — keyed object of `TrackingEntry` records; new jobs saved with status `"saved"`; re-encountered jobs have
+    `last_seen_at` and `updated_at` refreshed, all other fields (including user-set status) preserved; one disk write per
+    `upsertEntries()` call regardless of batch size
+  - `runs.jsonl` — append-only newline-delimited JSON log; one `BriefingRun` per line via `appendRun()`
+  - Constructor accepts `trackingPath`, `runsPath`, and `dryRun` overrides; defaults from `config.ts`
+  - `load()` returns empty store on a missing or corrupt file — no crash
+  - `ensureDir()` creates parent directory recursively on first writing
+  - `upsertEntries()` returns accurate `{ created, already_present }` counts even in dry-run mode
+  - `ScoredJob?` parameter on `upsertEntries()` reserved for Phase 8 score snapshot attachment
+- `src/tests/tracking.test.ts` — 23 unit tests; per-test isolated `tmpdir` directories; covers load, upsert (new + re-encounter),
+  disk writes, JSONL append, and dry-run suppression
+
+### Notes
+
+226 tests across 12 files, all passing. No new dependencies. The `TrackingRepository` is the last infrastructure module before the
+Phase 8 briefing orchestrator wires everything into the full pipeline.
+Dry-run mode is a first-class concern throughout: all writing paths are suppressed, all read and count paths remain fully functional, matching
+the Python careerclaw `--dry-run` behaviour.
+
+---
+
+## [0.6.0] — 2026-03-04
+
+### Added
+
+- `src/drafting.ts` — `draftOutreach(job, profile, matchedKeywords)`:
+  deterministic outreach email generator; subject line follows `Interest in {title} at {company}` format; body inserts experience
+  clause (years or "extensive experience" fallback), up to 3 matched keywords formatted as natural language, and 3 fixed
+  reliability/collaboration/instrumentation bullet highlights; word count 161–168 words depending on a keyword path, inside 150–250 word
+  spec; `llm_enhanced: false` always; `formatList()` helper for natural-language list formatting
+- `src/tests/drafting.test.ts` — 20 unit tests
+
+### Fixed
+
+- Deterministic template body was 127 words on the first pass — below the 150-word MVP spec floor; fixed by expanding opening and closing
+  paragraphs; both keyword and fallback paths re-verified at 161 and 168 words respectively
+
+### Notes
+
+203 tests across 11 files, all passing. No new dependencies. `llm_enhanced` is always false in this phase — LLM enhancement
+(Phase 7+) will set this flag to true when the Pro key is configured and the call succeeds. The deterministic template remains the permanent
+fallback for the Free tier and for LLM failure scenarios.
+
+---
+
+## [0.5.0] — 2026-03-04
+
+### Added
+
+- `src/requirements.ts` — `extractJobRequirements(job)`: tokenizes job title and description into a deduplicated keyword and phrase corpus for
+  use as the job-side input to gap analysis
+- `src/resume-intel.ts` — `buildResumeIntelligence(params)`:
+  section-aware keyword/phrase extraction across skills (weight 1.0), summary + target_roles (weight 0.8), and optional resume_text (weight
+  0.6); per-keyword weight is the max across sections; `impact_signals` are keywords with weight >= 0.8; `source` flag indicates which inputs
+  contributed; PR-E fix (skills injection) baked in from day one
+- `src/gap.ts` — `gapAnalysis(intel, job)`: weighted `fit_score` (sum of matched keyword_weights / job keyword count), `fit_score_unweighted`
+  (Jaccard), `signals` (resume ∩ job), `gaps` (job − resume), and top-5 `summary` for display
+- `JobRequirements`, `ResumeIntelligence`, `GapAnalysisResult` interfaces added to `src/models.ts`; `ResumeIntelligence` schema is
+  JSON-compatible with Python careerclaw output
+- `src/tests/resume-intel.test.ts` — 19 unit tests
+- `src/tests/gap.test.ts` — 16 unit tests
+
+### Fixed
+
+- Added `"am"` to `STOPWORDS` in `src/core/text-processing.ts` — missed from an initial set alongside `"is"`, `"are"`, `"was"`, `"were"`, `"be"`;
+  caught by resume-intel stopword filter test
+
+### Notes
+
+183 tests across 10 files, all passing. No new dependencies. The `fit_score` weighted formula is identical to the Python careerclaw
+implementation: skills listed in UserProfile. Skills receive weight 1.0 and will never appear as gaps. The practical fit_score ceiling against
+real job postings is ~50% due to company names and location tokens in the denominator.
+
+### Future Work
+
+- CorpusCache: Entropy-based token filtering (IDF) to suppress tokens that appear in >80% of fetched jobs. Gated behind corpus_size >= 50.
+  Planned for a future release after job tracking accumulates sufficient data.
+
+---
+
+## [0.4.0] — 2026-03-04
+
+### Added
+
+- `src/matching/scoring.ts` — four pure scoring functions:
+  `scoreKeyword()` (Jaccard token overlap, returns matched and gap keyword lists), `scoreExperience()` (clamped linear user/job years ratio),
+  `scoreSalary()` (proportional against a user minimum),
+  `scoreWorkMode()` (exact=1.0, hybrid=0.5 partial, mismatch=0.0);
+  `compositeScore()` with `WEIGHTS` (keyword=0.50, experience=0.20, salary=0.15, work_mode=0.15)
+- `src/matching/engine.ts` — `rankJobs(jobs, profile, topK)` scores all jobs, sorts descending by composite score, returns top-K `ScoredJob[]`
+  with full breakdown and keyword lists; scores rounded to 4 d.p.
+- `src/matching/index.ts` — barrel export for matching public API
+- `src/tests/matching.scoring.test.ts` — 36 unit tests
+- `src/tests/matching.engine.test.ts` — 10 end-to-end tests using real model types
+
+### Notes
+
+148 tests across 8 files, all passing. No new dependencies. Neutral score (0.5) is used for all null data cases, so missing job fields
+neither reward nor penalize the composite — consistent with Python careerclaw behavior. Gap keywords from `scoreKeyword()` feed directly
+into Phase 5 gap analysis.
+
+---
+
+## [0.3.0] — 2026-03-04
+
+### Added
+
+- `src/sources.ts` — source aggregation layer: `fetchAllJobs()` runs both
+  adapters concurrently with per-source error isolation; `deduplicate()`
+  removes duplicate `job_id` entries (first-seen wins); returns `FetchResult`
+  with job list, per-source counts, and error map for run instrumentation
+- `src/core/text-processing.ts` — shared text processing library:
+  `STOPWORDS` (English function words and full PR-E recruitment boilerplate set),
+  `SECTION_WEIGHTS` (skills=1.0, summary=0.8, experience=0.6, education=0.4),
+  `tokenize()`, `tokenizeUnique()`, `extractPhrases()`,
+  `extractPhrasesFromText()`, `tokenOverlap()`, `matchedTokens()`,
+  `gapTokens()`
+- `src/tests/text-processing.test.ts` — 34 unit tests
+- `src/tests/sources.test.ts` — 10 unit tests (ESM-safe adapter stubs via
+  `vi.doMock()` + `vi.resetModules()`; no network)
+
+### Notes
+
+102 tests across 6 files, all passing. No new production dependencies.
+`SECTION_WEIGHTS` is defined here and will be consumed by resume intelligence
+in Phase 5. The `FetchResult.errors` map feeds into `BriefingRun.sources`
+instrumentation in Phase 8.
+
+---
+
+## [0.2.0] — 2026-03-03
+
+### Added
+
+- `src/adapters/remoteok.ts` — RemoteOK RSS adapter; parses RSS XML into
+  `NormalizedJob[]`; `parseRss()` exported separately from `fetchRemoteOkJobs()`
+  so contract tests call pure parsing functions without network mocking
+- `src/adapters/hackernews.ts` — HN Firebase adapter; fetches "Who is Hiring?"
+  thread comments in parallel; `parseComment()` exported for offline testing;
+  handles deleted/dead items gracefully
+- `src/adapters/index.ts` — barrel export for all adapter public API
+- `src/tests/fixtures/remoteok.xml` — RSS fixture covering full fields, no-salary,
+  and k-suffix salary variants
+- `src/tests/fixtures/hn-thread.json` — HN thread fixture with `kids` array
+- `src/tests/fixtures/hn-comment-job.json` — HN job comment fixture (pipe-separated
+  header, HTML body, salary, experience years)
+- `src/tests/fixtures/hn-comment-deleted.json` — deleted comment fixture (adapter
+  must skip)
+- `src/tests/adapters.remoteok.test.ts` — 25 offline contract tests (title/company
+  splitting, salary parsing, work-mode inference, HTML stripping, `stableId`)
+- `src/tests/adapters.hackernews.test.ts` — 18 offline contract tests (header
+  parsing, timestamp conversion, HTML decoding, skip logic for deleted items)
+- `scripts/smoke_sources.ts` — live smoke test hitting real RemoteOK RSS and HN
+  Firebase APIs; run manually before releases with `npm run smoke`
+- `fast-xml-parser` added as a production dependency (RSS parsing)
+- `tsx` added as a dev dependency (runs a smoke script without a compiler step)
+
+### Changed
+
+- `stripHtml()` fixed: opening `<p>` tags now convert to `\n` (was `""`) so HN
+  comment header and body lines split correctly after HTML stripping
+- `README.md` — roadmap updated; Phase 2 marked complete; note updated to
+  reference v0.2.0
+- **Payment processor:** Pro license switched from Gumroad to **Polar.sh**
+  (`https://polar.sh/orestes-garcia-martinez/careerclaw-pro`); `CAREERCLAW_PRO_KEY`
+  env var name and SHA-256 cache behavior unchanged
+
+### Notes
+
+58 tests across 4 test files, all passing. No network calls in CI — all adapter
+tests use offline fixtures. Run `npm run smoke` manually before each release to
+validate live sources.
+
+
+## [0.1.0] — 2026-03-03
+
+### Added
+
+- Initial repository scaffold: `package.json`, `tsconfig.json`, `vitest.config.ts`
+- `src/models.ts` — canonical data schemas (`NormalizedJob`, `UserProfile`,
+  `TrackingEntry`, `BriefingRun`, `ScoredJob`, `OutreachDraft`); identical
+  JSON serialization format to Python careerclaw for cross-implementation
+  file compatibility
+- `src/config.ts` — centralised environment and source configuration
+  (runtime paths, HTTP defaults, RemoteOK RSS URL, HN thread ID, LLM and
+  license env vars)
+- `SKILL.md` — OpenClaw skill definition with Node-native self-healing
+  install check (`npm install -g careerclaw-js`)
+- `CHANGELOG.md`
+- Unit tests for `models.ts` and `config.ts` (Vitest)
+
+### Notes
+
+This release establishes the Phase 1 foundation types. No adapters,
+matching, or CLI are included yet — those follow in Phases 2–8 per the
+Node Migration Decision (ADR, March 2026).
+
+[Unreleased]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.8.0...HEAD
+[0.8.1]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.8.0...v0.8.1
+[0.8.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.7.0...v0.8.0
+[0.7.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.6.0...v0.7.0
+[0.6.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.5.0...v0.6.0
+[0.5.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.4.0...v0.5.0
+[0.4.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.3.0...v0.4.0
+[0.3.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.2.0...v0.3.0
+[0.2.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/compare/v0.1.0...v0.2.0
 [0.1.0]: https://github.com/orestes-garcia-martinez/careerclaw-js/releases/tag/v0.1.0