…es, architecture

Refresh the line-count claim (~2,600 src + ~850 test lines after the
recent fixes), document the min/dev branch model, add the
CORECODER_TIMEOUT_MS env var, describe the fixed-overhead calibration
accurately, and list scripts/ in the layout.

Drop the npm install command (package is not published), mark the
Ollama quick start as needing a dummy key, scope the .env lookup claim
to the home directory, soften the design-commentary claim, and reword
the branch descriptions (min: kept as small as possible; dev:
enhancements no longer bound by the minimal bar).

Overhaul the README: from one sentence to something you can actually use

The --help usage string (and the README block mirroring it) claimed
$OPENAI_API_KEY is the default key source, but config.ts resolves
$CORECODER_API_KEY first. State the real precedence in both places.

package.json has declared MIT since the initial commit, but no LICENSE
file existed, so there was no actual grant text and GitHub reported the
repository as unlicensed.

Fix --api-key help text and ship the MIT LICENSE file

P0 (data corruption / dead-on-install / documented-usage traps):
- edit_file: replace via a function so $$/$&/$' in new_string are
  written literally instead of corrupting the file
- CLI: realpath argv[1] before comparing to import.meta.url — npm bin
  symlinks made the installed command silently do nothing
- Agent.chat: finally-backfill tool replies when a consumer abandons
  the generator (for-await break poisoned the history permanently)
- SSE body phase: abort now cancels the reader (unblocks reads and
  closes the connection) and each read races an idle timer, so a
  stalled stream errors instead of hanging the turn forever

P1 (hangs / silently wrong):
- bash kill: SIGTERM now escalates to SIGKILL after a 5s grace, so a
  TERM-trapping child can't wedge the REPL; timer cleared on exit
- tools: makeAllTools() factory — a second Agent no longer re-points
  the first Agent's shared SubAgentTool at itself
- context: compression LLM calls take the turn's AbortSignal
- llm: tool-call arguments parsing coerces null/array/scalar to {}
- context: layer-0 truncation for a single message that alone exceeds
  the window (was an unrecoverable 413 loop)
- glob: strip leading ./ and support [ch]/[!x] character classes
  (both silently matched nothing)
- cwd: tracked cwd moved to paths.ts so every tool resolves relative
  paths against it, and cd parsing follows ;/|| chains too

Adds 11 regression tests (54 total pass).

…nes)

- Revert the cross-tool cwd unification and ;/|| parsing (magic state,
  out of min's scope — the split-brain stays a documented limitation)
- Replace silent oversized-message truncation with an up-front refusal
  in Agent.chat (the truncation math ignored the fixed overhead and
  silently altered user input; refusing is honest and simpler)
- Kill sequence now starts at most once (a second trigger orphaned the
  first SIGKILL escalation timer, which settle() could never clear)
- rounds() closes the inner LLM generator in a finally, so abandoning
  a turn during text streaming cancels the SSE reader and releases the
  connection (hand-driven .next() gets no auto-close from yield*)
- isDirectRun's catch falls back to a plain compare instead of
  silently refusing to start
- Keep the [ch]/[!x] character classes: verified native in Python
  pathlib.glob — this is port parity, not added syntax

Tests: drop the cwd case, add oversized-refusal and
abandon-during-text-streaming-closes-connection. 54/54 pass.

…asses

- rounds() awaits stream.return() and sseEvents awaits reader.cancel():
  for-await's break waits on the return() chain, so teardown now
  completes before break returns, and a throwing inner finally surfaces
  as a handled rejection instead of an unhandled one. The regression
  test asserts closure immediately after break, no sleep.
- Character classes now follow Python fnmatch semantics on the edges:
  ^ is a literal member (only ! negates), ] right after [ or [! is a
  literal member, and an invalid range like [z-a] matches nothing
  instead of throwing. The doc comment states the supported subset
  explicitly — no claim of a full glob dialect.

…onesty

- kill goes straight to SIGKILL on the whole process group: Python's
  subprocess.run timeout behavior (kill outright), hardened from the
  direct child to the group. The TERM->timer->KILL machinery is gone.
- remove the oversized-input pre-check: it judged user input by an
  estimate and could refuse legitimate messages; the provider's own
  context error is the honest answer.
- /tokens prints 'unknown' when the provider has never reported usage,
  instead of dressing unknown up as 0 (usageSeen on LLMClient).
- write the token-accounting principle into context.ts, document the
  no-usage degradation on observe(), compress the sseEvents comments.

Prices rot within months, the estimate multiplies a guessed token
count by a guessed rate, and the table teaches nothing architectural.
Token accounting stays — that is real provider data; only the money
guess goes. Removing estimatedCost from LLMClient is a breaking
change for library consumers, acceptable pre-1.0.

Fix eleven correctness bugs from the source review

…rc / 1,104 tests)

Remove cost estimation (PRICING table and estimatedCost)

Four follow-ups from the round-4 review of the merged fixes:

- Calibration: measure() is now messageEstimate x ratio + fixedEstimate
  (and observe derives the ratio from real minus fixed). The blended
  formula multiplied the English fixed overhead by a CJK conversation's
  rate, overstating a compressed history by up to 2x the overhead
  (simulated: 565 vs a true ~232). Adds the mixed-rate regression test
  the old suite lacked.
- Usage honesty: usageSeen only flips on actual numbers (a usage object
  with null fields is not data), and a new optional usageMissed flags
  responses that carried none, so /tokens now has three states:
  unknown / reported-but-incomplete / reported. The label says
  'Reported tokens' — no more claiming a complete total.
- Tool exports: drop ALL_TOOLS and getTool (zero internal or test
  usage); the shared SubAgentTool instance they exposed reintroduced
  the cross-agent parent-stealing bug for library consumers. Only the
  makeAllTools factory remains.
- npm test: clean runs first (portable node -e rm, no more rm -rf), so
  compiled remnants of deleted tests can no longer execute. Verified by
  planting a failing ghost test in dist/ — it no longer runs.

Replace the ratio-calibration subsystem with a stateless heuristic and
a strict two-lane rule — the estimate drives pre-call decisions and is
never calibrated; provider usage is post-hoc accounting and never
feeds the estimate.

- approxTokens = ceil(utf8Bytes / 3): ASCII lands at ~3 chars/token,
  CJK at ~1 char/token (3 bytes/char), and byte counting sidesteps
  UTF-16 surrogate pitfalls — no Unicode tables, no state.
- delete ratio and observe() (and the agent call site); measure() is
  the plain estimate plus the fixed prompt/tool-schema overhead.
- usage accumulates only when a chunk carries BOTH prompt and
  completion numbers; the incomplete mark now lands in a finally, so
  an aborted or torn-down stream flags the totals too (post-loop code
  never runs on teardown).
- /tokens: 'Reported tokens: P / C (incomplete)' — three states,
  no total claim.
- README: the three calibration passages now describe the byte
  estimate and the two-lane rule.
- tests: calibration cases replaced by exact deterministic byte-math
  assertions; new partial-usage and aborted-stream-incomplete cases.

A stream could deliver the complete usage chunk and then die before
[DONE]: the finally saw complete numbers (so no incomplete mark) while
the accumulation after the loop was skipped by the teardown — a ledger
claiming completeness with the numbers missing. Accumulation now lives
in the finally: book the numbers when complete, mark incomplete when
not, one lifecycle event either way.

Also: agent.ts comment no longer mentions calibration, and the test
file is renamed context-estimate.test.ts to match what it tests.

Shrink token accounting: static UTF-8-byte estimate, complete-usage-only bookkeeping, tool exports, stale tests