# REPORT — harness skeleton (build phase 2) — 2026-08-28

## Works

- **Config**: `harness.toml` via a stdlib-only TOML-subset parser (tables,
  bare keys incl. hyphens, strings/ints/bools, single- and multi-line
  arrays; trailing commas tolerated; anything richer fails loudly with
  line numbers). Defaults + validation; secrets are refs only
  (`env:NAME` / `file:PATH` / `literal:VALUE`; `bw:` reserved) and redacted
  from every error path. `harness.toml.example` is tracked and load-tested
  so it can never rot; real configs are gitignored.
- **Model routing v0**: `[models]` tier map + `[models.classes]` class map;
  requests go out with the CONCRETE model resolved from the map
  (mopac-study→glm-4.7-flash, mopac-code→glm-5.2, mopac-review→glm-5-turbo,
  mopac-primary→glm-5.3, + mopac-vision→glm-4.6v). No heuristic code;
  unknown class = hard error naming the config section.
- **Conductor single-shot**: `harness once` = intake → routing → bounded
  turn → REPORT, then exit 0. Chainable, no daemon. `--dry-run` = intake +
  plan (resolved model, tool bounds) with ZERO LLM calls (test-asserted).
  `--task-id` filter for chaining. Exit codes: 0 ok/no-tasks, 1
  config/usage, 2 intake, 4 llm/turn.
- **Intake**: Redmine `/issues.json` scope query (raw filter params or
  saved query id), task class from a configurable custom field with
  default-class fallback; `--demo` builds the issue from `[demo]` (no
  Redmine needed).
- **Bounded turn**: OpenAI-compatible chat via LiteLLM (stdlib http;
  base_url ± `/v1` normalized; Bearer auth; retry/backoff on 429/5xx/
  transport; usage accounting), tool-calling loop capped at `max_rounds`;
  gate denials feed back to the model as tool results and are counted, not
  fatal. On mid-turn LLM failure after content exists, a partial REPORT is
  still written with the error as stop reason.
- **Exec tool**: allow-listed bash. maki-derived scope semantics without
  tree-sitter: compound commands split segment-by-segment (`&&`/`||`/`;`/
  `|`, quote-aware), `cmd *` word-boundary, `pfx*` raw prefix, `pfx/**`
  path prefix, `*` universal; deny beats allow; `$()`/backticks/subshells
  always denied (headless has no prompt channel). Per-command timeout with
  process-group cleanup on unix; output truncation.
- **Writeback**: `REPORT-<vertical>-<task>-<ts>.md` + `REPORT-latest.md`
  (atomic tmp+rename) with model/tier/class/tokens/rounds/denied/stop
  telemetry per the DESIGN auditability bar.
- **Tests**: table-driven, stdlib `testing` only — TOML subset (valid doc
  + 9 error cases), config defaults/validation/key refs (incl. leak check),
  routing decisions, scope matcher + gate (14 cases) + exec (timeout,
  truncation, deny), writeback, Redmine intake (httptest), LLM client
  (auth/model/retry/4xx/empty-choices), and loop end-to-end against a
  scripted fake OpenAI server (demo turn, dry-run zero-call, tool
  round-trip with message-shape assertions, denial counting, round limit,
  error-class mapping). `go build ./...`, `go vet ./...`, `go test ./...`
  all clean on go1.26.7; every intermediate commit builds standalone.

## MVP demo bar status

`harness once --demo` is wired end-to-end: prompt "tell me about yourself"
→ LiteLLM (glm-5.3 via mopac-primary) → GLM self-description → REPORT.
Proven against a scripted fake in tests. The live proxy
(http://192.168.3.78:4001) is reachable from this host and answers 401
without a key, so the live run needs exactly one thing:

    export HARNESS_LITELLM_KEY=<vertical virtual key>
    ./bin/harness once --demo

## Stubbed / known gaps

- Redmine issue-note writeback (SoR note after the REPORT) — file only.
- Budget/semaphore GATE (LiteLLM spend APIs, class-aware slots) and cost
  in REPORT (tokens only today).
- `bw:` key refs error until the bitwarden wrapper (phase 3).
- Streaming + turn resume on truncation — retry is request-level today.
- Session persistence + read-time repair (crush notes §2) — not started.
- Write confinement to declared roots + symlink-aware canonicalization;
  redirects currently match by segment text (phase 3 permission layer).
- Task selection is first-in-scope; no P1-first ordering or status
  transitions, so chained `once` re-picks the same issue (use --task-id).
- Local inbox intake (DESIGN core-loop step 1, second half) — not started.

## Next chunk (phase 3)

1. bitwarden-go wrapper + `bw:` refs (unblocks secret posture).
2. Full permission layer: tree-sitter bash scopes, write roots, symlink
   checks, per-vertical allow/deny presets.
3. Redmine note writeback + status transition so chaining advances scope.
4. Budget gate via LiteLLM spend APIs; cost line in REPORT.
5. Streaming with truncation retry + turn resume.