Status and limits

Everything below is built and tested; less of it has run against real daily use. The one result that matters, a skill kept from daily use, has not been earned yet.

Built

Phase	Built
1	the interactive loop
1.5	create and improve, with the mechanical split and the forge import
2	batch mode with worktree isolation and the next-run keep rule; context-hook unit retrieval; tool-output truncation and stuck detection as entries under the same counters (three built-in hook entries that wait for your approval)
3	team and fork shipping, work-to-home export, stale on drift, split per family, transfer suggestions and shared promotion
—	omp profile resolution (OMP_PROFILE / PI_PROFILE / PI_CONFIG_DIR / PI_CODING_AGENT_DIR), multi-profile history import with --skip-profile and policy.json skip_profiles
—	the lake, L1–L4 (ADR 001)
—	the profile: rematch profile init|adopt|sync|status|install-launchd

Run live

• The loop, against omp 18.12.6 on a scratch repository: rpc-mode sessions with real redirects, gemini-4-argon as editor and classifier, the entry loaded in the next session and counted.
• Also against omp: the forge import and the mechanical split of a 31 KB skill resolved as a package in a live session, one detached batch cycle with a script verifier, one team ship, and /rematch in the TUI.
• The lake, on both of the author’s machines on 2026-10-26: 1,178 journal rows and 326 session event files exported in 2.9 s, all keys indexed with Bellwether 0.1.0 in 0.8 s, report from the lake in 57 ms, why --asof giving the pre- and post-transition status of one entry, and the second machine’s lake reading this one’s history after lake import (ADR 001, “Live run”).
• The profile, on 2026-10-26 between two Macs over a private GitHub repo (the profile). It is also covered by crates/rematch/tests/e2e_profile.rs on two temporary machines sharing a local bare remote.
• Claude Code, with claude -p --model haiku (Claude Code 2.1.285) against a temporary root: every trace record, a negation pre-labelled, flagged and counted once; reflect --pending started by the next SessionStart; a prompt queued during a Bash call traced as interrupt; the digest with approval commands; a provisional agent copied, listed and counted.
• Not yet observed live in Claude Code: the digest’s display in an interactive session, an Esc abort (covered by an end-to-end test on the transcript format), and a session reflected by a real editor model.

Pi and OpenCode runtime checks

Both adapters are exercised inside their real pinned host binaries with isolated state and a local scripted model. These checks test integration, not cloud inference or agent-quality improvements.

Host	Exercised
Pi 1.4.0	Native resource discovery, input/tool/error traces, user approval and approved blocking, resume/reload, off/on, and an ordinary prompt after an abort
OpenCode 1.19.4	Plugin and native skill loading, success/error traces, approved blocking and retrieval, idle continuation, process disposal/crash recovery, resumed segments and counter idempotence

OpenCode’s truncation policy and stuck-repeat edges have deterministic tests. Its TUI toast display has not been exercised interactively.

Not yet earned

• A kept skill from daily use. The three built-in hooks reached kept (9/6/5 matched sessions, 0 recurrence, 0 harm) during the build itself; no skill has yet. This is the kill criterion.
• The built-in context entries in real use. All three were approved and fired in one live omp session (a retrieval record for a bellwether-dev unit; the stuck detector blocking a 4th identical call). Truncation is observable only in the outgoing request and is covered by the extension tests. None has been counted across real sessions yet.
• A batch cycle where a candidate beats the reference. The one live cycle had a reference that already passed, so both editors chose KEEP_BASE; the fake-omp tests cover the other branches.
• A merged team PR and a shared promotion from real branches. Both are exercised on scratch repositories only.
• Linux daily use. Native Ubuntu CI runs the Rust and omp-extension tests; the model-backed examples above were run on macOS arm64.
• Windows daily use. The omp path has PowerShell installers, native process-tree cancellation and native CI, including an isolated installer/CLI smoke. Enable Developer Mode for symlinks. POSIX sh fixtures run on Unix only; non-omp host runtime parity is not yet certified on Windows.
• Embeddings for the related-thing lookup. Exact signature and MinHash are implemented; embeddings would need credentials the binary does not hold.

Open questions from the spec are at the end of docs/SPEC.md.

Windows .cmd/.bat omp launchers receive model prompts through stdin. This avoids shell interpolation, embedded-newline rejection and the command-line length limit. The prompt still goes through omp; no provider credentials are added to Rematch.

Trace filenames on Windows escape reserved characters such as :. Logical session ids in records do not change. New lake-event filenames are portable on every platform, and older Unix exports remain readable.

Claude Code limits

Each limit is structural; the reasons are in DESIGN §18.1, “What Claude Code cannot do”.

Limit	Why
no truncate hooks	no hook rewrites earlier tool results or the outgoing request; PostToolUse replaces a result only as it arrives
retrieve once per prompt, and its units stay in the conversation	there is no per-request hook (omp re-sends one non-persisted message per request); an unchanged unit set is not injected again
skill loading is not per process	the shared dir serves every session and harness; a concurrent session sees new links only after its next re-scan. Its trace counts only what matched its own cwd.
no per-turn event	turn_end records are recovered at Stop from the transcript tail, which can lag; model fallbacks inside a run show only as turn-to-turn changes
Esc fires no hook	the abort is read from the transcript at the next prompt, and that prompt is not a takeover
stuck sees completed calls only	parallel identical calls in flight are not counted; subagents get block rules but no stuck detection
no approval dialog, no rematch tool for the model	hooks cannot open dialogs; approvals are commands you run
reflection needs omp	the model calls go through omp; without it, run rematch reflect --pending from a shell
agent copies refresh at session start only	a status change reaches Claude Code at the next start; omp-only frontmatter is dropped from the copy
a session killed before SessionEnd stays unfinished	count skips it, the same as an omp crash
imported sessions carry no loaded_entries	transcripts do not record what the plugin linked
batch mode is omp-only	the batch worker is omp-typed; the plan for another harness is DESIGN §18.2.4

Codex limits

Post-hoc reflection, post-hoc counting and skills are the whole adapter.

Limit	Detail
no live tracing	rematch has no code inside Codex; a session becomes evidence only when you import its rollout
no blocking hooks	block/stuck verifiers never run; a hook entry’s router is listed by Codex but never counted as loaded
no retrieval, no truncation	both need a pre-request handler
no per-session gating	Codex loads every skill under its skill roots in every session; a provisional skill stays visible everywhere until kept or retired, but counts only where its predicate hits
loaded entries are reconstructed	resolved at import from the rollout’s skill list and the files as they are then
no subagents	kind: agent entries have no Codex form; scripts in a linked skill stay non-executable until approved
no batch mode yet	rematch run drives omp; the Codex plan is DESIGN §18.2.4

Pi and OpenCode limits

• Pi: no native subagent definitions; Rematch agent entries are not delivered there.
• OpenCode: no generated agent-profile renderer, live tracing of child sessions, or approval dialog. Child sessions can be imported as history.
• OpenCode: retrieval and truncation use experimental request-transform hooks in 1.19.4. OpenCode 2 is not a tested target.
• Both: reflection and batch solver calls still use omp. The adapters are installed from a checkout, not a registry package.

Full lifecycle and API evidence: Pi adapter and OpenCode adapter.

How it is checked

Every change goes through:

cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
bun build extension/index.ts --no-bundle --outdir /tmp/rematch-ext-check
bun test extension                            # hook runtime: mode helpers + the factory against a fake host

The end-to-end tests run the built binary against a temporary root:

• a synthetic trace with a redirect is reflected through a scripted editor, staged, matched, counted through four clean sessions and kept;
• the lint-fail script exercises exactly one retry; a wrong-model reply fails closed;
• import --history runs against anonymized goldens;
• improve refuses, splits, stages a sibling and refuses a pinned skill;
• the batch worker completes a k=1 cycle with a fake omp and keeps the winner on the next run;
• ship creates the branch and commit; export refuses at work by default.