A structured model function such as the tutor's ask_model now works
as `ask` in check, looks_like_code and ask_models (via: "model").
Raises fall through to Ollama, then fail open as before. v0.1.1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWYGW5NZL5uVpnqmNKo
Standalone utils subproject (pyproject, pre-commit, CI, install.sh,
README with the library API). It no longer imports leetcode_guard:
guard_bridge resolves the three read-only calls on demand.
Q5: SQL problems skip req: and ask for an OUTPUT SHAPE (columns,
rows, grouping) before code; algorithm problems keep req:.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWYGW5NZL5uVpnqmNKo
tool/coach_live_check.sh starts the real Relay (lane, backend, HTTP API,
phone mirror) on a free loopback port, never 8777, with claude, Firebase,
LeetCode, Ollama and Anki faked and every file under a temp dir
(_live_relay.py), then drives it through WrapperServer's /coach/* proxy
with WrapperCoachTransport (coach_live_check.dart):
- flow: start, step-1 partial then pass, a chat turn, code refused before
step 2, three step-2 fails via toolbox, a force, an unknown and an
over-long pick (bad_request), an also_accepted pass, run_code, end,
not_live after it; 404 for an unknown rid.
- slow: a 35 s coach turn through a 30 s proxy timeout answers, with
heartbeat.busy naming the rid and every poll under 2 s.
The flow's raw answers are committed as test/fixtures/coach_relay/ and
parsed by relay_fixtures_test.dart, so the models are checked against
what the relay sends, not what the fake's author believed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
phase.step2_attempts counted every step-2 grade with grade "fail",
and a force repeats the last try's grade, so three named-a-tool fails
plus "Continue anyway" showed attempts: 4 in the state. Measured
against the live relay through the app's desktop path; the contract
(and the snapshot's _coach_grades) already exclude forces.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The RTDB mirror wrote synchronously on the lane's one worker, so a hanging
Firebase (15 s connect timeouts) would have stalled every desktop request.
PhoneLink now only enqueues; relay.outbox writes on its own thread in the
contract's order (chat reset/lines, state, heartbeat, then the result),
keeps only the latest state and heartbeat, never drops a result or a chat
line, and retries a failed write first, with backoff.
The whole-request 280 s cap is enforced again: Coach.handle sets a
deadline and a coach turn gets what is left of it (at most 240 s), so an
ambiguous message's code check plus the turn cannot outlive the phone's
300 s "No reply" wait. The contract's caps line is back to that wording.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
leetcode_coach.relay implements DOCS-coach-relay.md: one paced serial lane
(1 start / 3 s, 5 queued, done-map re-answers a rid), the loopback API on
127.0.0.1:8777 (Host check, JSON only, 16 KiB), the RTDB request/result/
state/chat/heartbeat mirror over SSE with 60 s rescans and 24 h cleanup, the
five-step state machine derived from the session log, graders (steps 1/5
haiku -> Ollama, fail open "unverified"; step 2 through PhaseGate.screen;
step 3 mechanical against answer_key.json), the CLI's sonnet chat with the
log in the child env and the local hand-over, run-code on node --permission,
and step 4 through the harvester's read-only fetch_recent_ac once a minute
(three-valued: unknown is never "not yet").
New log events session_start/grade/chat/session_end land in session_log,
_coach_log and _well_formed together; the snapshot rows gain grades and
accepted (_coach_grades). A forced step-2 grade opens code (phase.req_state).
leetcode-coach-relay.service + an install.sh step (not run).
Contract fixes in the same commit: step 3 counts also_accepted and combined
entries' parts, step-2 grade events and the force event shape, next_step,
per-call model caps instead of a 280 s watchdog, heartbeat on busy changes,
the statement fetched in-process, enable + restart on reinstall.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
The exporter used to publish coach: null before the first session, so the
tab printed "unknown" everywhere with no way forward. Now a missing or empty
sessions dir exports empty problems/re-tests, every cluster at zero with its
problems (slug, number, title, URL) and next, and a how_to block (commands,
the 4-step method, first command). null stays for a real failure.
how_to lives once in leetcode_coach/how_to.py; DOCS-how-to.md quotes it and a
test fails on drift. The app's Coach tab shows a How to start card (monospace
command box with copy), one card per cluster with tappable problems, re-tests
with their command, and "unknown" + why only for a null section. Also fixes
the done/solved key mismatch (the app read a key no exporter wrote).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
The incomplete-list note shows meanings before code, so "a ^ x ^ x = a"
or "x & -x" handed over the answer. Rewrite the six leaking meanings plus
two that said "sorted", and pin it with
test_no_property_meaning_names_an_entry.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
The fetched statement carries LeetCode's topic tags and site hints ("Hash
Table", "Two Pointers"), so exempting it pre-cleared the answer. Some real
toolbox meanings illustrate with the operation (self-inverse: "a ^ x ^ x"),
so the incomplete-list note shows those properties by name only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
The gates read the toolbox through toolbox.load_safe (typed, fail-open)
beside main's raw load(); the session log follows LEETCODE_COACH_SESSIONS_DIR,
stamps LEETCODE_COACH_CLUSTER, and the card builder logs postmortem on
delivery. leetcode report dispatches beside cluster_cli; client notes move
to DOCS-clients.md to keep SKILL.md under the cap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
"merge intervals" is #56's own title, so the leak bounce would have bounced
the coach's restatement of it; "dictionary" was a bare common word. A test
now pins that no entry name or alias equals any clusters.json title.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
- toolbox.json: 20 entries over a fixed 40-property vocabulary; no entry's
property set nests in another's, so every property card has one answer.
- clusters.json: decision 10's six clusters, ids plus slugs.
- progress.py (stdlib only): solved = code + postmortem, re-tests due at
+3/+10 days, cluster progress, all derived from the session logs.
- leetcode due / --cluster <name> / --retest <slug> (cluster_cli.py): due
re-tests of the cluster first, then the next problem with no code event,
skipping LeetCode-Accepted ones (live check fails open). A cold re-test
logs `retest` only once the session reached a post-mortem.
- Post-mortem cards are property -> operation only, generated by
toolbox_cards (shared slug `toolbox`); `add` appends an entry only after
kuhy confirms its properties. The local client accepts {"toolbox": [...]}.
- _coach_log publishes retests_due and clusters from the coach.
- requirements-sync treats sibling repo packages as first-party.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
The REPL passes os.environ to `claude -p`, so a PYTHONPATH set by the
launcher reached every hook and every Bash call of the coached session. The
old launcher did not do that. The launcher now runs `/usr/bin/python3 -P -m
leetcode_coach.leetcode_repl` and relies on install.sh's editable install,
the same way the Stop-hook shim does. A test now pins that the launcher
sets no PYTHONPATH.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
Phase A of TODO-coach.md (decisions 1-3). The skill code is now a package here
with the same behaviour as before. Nothing new is added. ~/.claude/skills/leetcode
and ~/.local/bin/leetcode become symlinks into the package, created by
scripts/link_leetcode_coach.sh, which install.sh runs.
- Package imports replace the bare sibling imports. To stay under the 250-line
cap, the Stop-hook detector is split into gate_detector.py and stop_gate.py,
the card rules move into card_rules.py, and the prompt text moves into
local_prompt.py.
- local_coach uses gate_detector directly. It no longer loads the hook file
with importlib.
- make_cards runs on the system python: a genanki pin (pyproject and
requirements) replaces the `uv run --script` header. The note guid is
passed to genanki.Note explicitly, so CardNote is no longer subclassed;
the guid values are unchanged.
- stop_gate also accepts the package path as a skill-load marker, because the
harness may print the resolved symlink path.
- Every swallowing except now logs a warning. A NullHandler on the package
logger keeps stdout/stderr the same as before.
- Modules whose prints are their interface (REPL, local session, card builder,
fetcher, eval, Stop hook) get T201 per-file ignores, and their tests assert
on stdout.
- hook_shims/ holds the 3-line ~/.claude/hooks shim. It is excluded from
ruff, mypy, pylint and coverage, and a test runs it end to end.
- Pre-commit: a requirements-sync entry for each package, leetcode_coach and
genanki added to the pylint hook, and the name-tests-test exclude widened.
coverage and ruff are bumped to the latest stable releases because the
dependency-freshness gate requires it.
Verification: 111 coach tests pass with 100% branch coverage. ruff, mypy
strict and pylint are clean (10.00). On the 370 assistant replies in 39 real
/leetcode transcripts, the old hook and stop_gate give the same verdict
(0 mismatches, 20 blocked). fetch_problem output (live two-sum) and
make_cards --dry-run output are byte-identical to the old skill, and so are
the note guids.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PjvF5DZG9JduSBYKxqyysi
