v0.7.0

• feat(adapters): the self-host tier is real — a local adapter for any
  OpenAI-compatible server (Ollama, vLLM, LM Studio, SGLang, DwarfStar/DS4):
  user-declared capabilities and context window (the router's capability floor —
  never guessed), $0 default pricing (overridable, never invented), an explicit
  note when a server reports no token usage, and byte-exact tool-call ID replay
  promoted into the shared conformance battery for every adapter. New
  glam run --local / routing.localOnly fail loud rather than silently
  falling back to a hosted provider. Catalog gains ollama/vllm/lmstudio rows
  plus dwarfstar-ds4 and ornith-1.0-9b/35b with honest beta/unverified flags.
  Live-verified against a real Ollama daemon: keyless glam run tool
  round-trip at cost: $0.000000. (#25)

v0.6.0

• feat(proxy): glam serve — the router-as-proxy gateway. A local
  Anthropic-Messages + OpenAI-chat-completions endpoint that puts glamfire's meter,
  router, hard budget stops, and owned usage ledger under agents you already run
  (Claude Code, opencode, Cursor, any OpenAI SDK). Streaming both dialects,
  tool-call IDs round-trip verbatim, bearer auth always required, loopback by
  default. Per-client budgets reject over-budget requests with a provider-shaped
  error before any provider is called. glam usage gains a by-client breakdown.
  Live-verified: real Claude Code headless through the proxy to GLM 5.2 on
  Fireworks — 2 tasks with real tool calls metered at $0.0423 vs a $0.303
  Claude-pricing estimate for the same session. See docs/PROXY.md. (#41)
• feat(brain): the markdown is the brain; SQLite is a rebuildable index. Every
  record is a readable markdown file with YAML frontmatter under brain/
  (sources/facts/notes/pointers/episodes + generated INDEX.md and log.md):
  truth: source|summary, sharing: personal|team, derived_from with per-link
  content hashes. New glam brain add|list|query|sync|lint|rebuild. Content-hash
  sync (file-wins for sources; conflicts surfaced, never merged silently). The
  tested invariant: delete the .sqlite, rebuild losslessly from markdown. (#36)

v0.5.0

• feat(engine/brain): memory is in the loop — every glam run recalls relevant
  records from the project's local brain store (.glam/brain.db, hybrid retrieval,
  hard token cap, full provenance in the packed context) and writes a structured
  episode (task, outcome, decisions, files touched, models, real cost) back after
  the run. Header/receipt show what was recalled; --json gets a memory object.
  Controls: --no-memory, [memory] in glam.toml, GLAM_MEMORY=false. The
  self-contained npm bundle (no native store) says so honestly and runs without
  memory — it never fakes recall. Verified live: a teach run then a recall run
  against real Fireworks/GLM; a --no-memory control run could not answer. (#27)
• feat(scripts): adopt-claude-code.mjs — brownfield seed of glam adopt
  (#29): discovers existing Claude Code state on this machine (global + project
  CLAUDE.md, project memory, auto-memory) and imports it into the project brain,
  idempotently, so runs recall it from the first task. Claude Code files are
  read-only inputs — never modified.
• feat(cli): BEHAVIOR CHANGE — a run stopped by a budget/step/token ceiling now
  exits 3 instead of 0, so scripts and CI can tell a budget stop from done
  without parsing output. The documented, stable scheme (glam help /
  glam run --help): 0 done · 1 error · 2 usage error · 3 budget/step ceiling ·
  130 interrupted (128+SIGINT). (#23)
• fix(cli): the run header now shows the serving provider and model family
  (e.g. provider: fireworks model: deepseek-v4-flash (…)) instead of leaking the
  shared adapter's internal id (adapter: fireworks-glm) on DeepSeek runs; adapters
  now declare a stable provider id alongside their adapter id. (#24)
• fix(cli): the recorded to … ledger line and the monthly-budget alert passed
  the useColor function (always truthy) instead of the resolved boolean, forcing
  raw ANSI codes into piped glam run output. Found while verifying #23.

v0.4.1

• fix(scripts): version.mjs main-module guard used naive file:// string
  concatenation, which never matches on Windows — node scripts/version.mjs printed
  nothing there (surfaced by the new doctor install-check test on Windows CI). Now
  compares via pathToFileURL.

v0.4.0

• feat(usage): local usage ledger (~/.glam/usage.jsonl) — every glam run records
  model, provider, tokens, real cost, duration, and per-model escalation splits; new
  glam usage command (totals, by day/model/provider, --since, --json, budget bar);
  opt-in monthly budget alerts ([usage] monthlyBudgetUsd / warnAtPct).
• feat(models): evergreen model/provider landscape — new glam models command over a
  built-in catalog (13 entries, live-verified prices with as-of dates and sources);
  --refresh pulls current provider prices, reports price drops, caches honestly;
  adapters price through the catalog so route/run/models can never drift.
• feat(adapters): DeepSeek V4 support — deepseek-v4-pro + deepseek-v4-flash on
  Fireworks (FP8, 1M context, live-verified incl. parallel tool calls, seed, and prompt
  caching) and DeepSeek-V4-Pro on Together (live pending key); adapter conformance
  63/63; Fireworks adapter now fails loud on unknown model ids.
• feat(cli): hardening/UX sweep — honest glam doctor inside compiled binaries
  (root-caused /$bunfs detection), real Ctrl-C interrupt (aborts in-flight provider
  request, honest partial cost, exit 130), numeric option validation, did-you-mean
  suggestions, NO_COLOR/FORCE_COLOR policy, EPIPE-safe piping, 2.3× faster startup.
• docs: context-wars messaging (README hero, WHY-WE-WIN), research briefs 24–25
  (creator-thesis update; provider/model landscape 2026-07 with cited prices).

v0.3.0

Full Changelog: v0.2.4...v0.3.0

v0.2.4

Full Changelog: v0.2.3...v0.2.4

v0.2.3

Full Changelog: v0.2.2...v0.2.3

v0.2.2

Full Changelog: v0.2.1...v0.2.2

v0.2.1

Full Changelog: v0.2.0...v0.2.1