20 lines
4.8 KiB
Plaintext
20 lines
4.8 KiB
Plaintext
Tickets confirmed. Now researching the z.ai usage API surface.
|
||
|
||
ERROR
|
||
|
||
Agent processing failed: failed to start agent processing stream: failed to fetch URL: Post
|
||
"https://sourcegraph.com/.api/graphql": context deadline exceeded (Client.Timeout exceeded while awaiting headers).
|
||
|
||
Codebase understood. Now researching the z.ai usage endpoint.Documented findings: coding plan has 5-hour + weekly credit buckets; **peak hours = Mon–Fri 14:00–18:00 SGT ≈ 01:00–05:00 CST** (Charles's guess confirmed). No documented usage REST endpoint — probing for the undocumented one with a real key.Gateway responds under `/api/coding/*` but no public usage route. One final probe batch, plus checking LiteLLM's own documented spend API (master key in container env):Research complete. Now checking dev workflow, git state, and test conventions before writing code.All context gathered. Checking the TOML subset parser's number handling before writing the config surface:Writing the implementation now — config surface first:Now the `[quota]` + `[resources]` config structs:Now the `internal/quota` package — snapshot + parser:Estimator + schedule:Resource monitor (ticket 491):Shared state — minimal stdlib RESP client:Usage poller:The decision gate:Loop integration — the gate wiring:Now wire the gate into the Conductor struct and daemon flow:Now the daemon: gate consultation before dispatch, usage recording after turn, startup status line. Also state.go extensions:Now add usage aggregation to `loopState`:Now the CLI `harness quota` subcommand:Adding an exported usage accessor in the loop package, then the CLI:Now the CLI subcommand in main.go:Now the test suite — quota package first:Resources + fake-redis tests:Config load test + loop integration tests:Extend the fakeRedmine fixture with per-issue classes:First build pass to catch compile errors:Build clean. Running the full test suite:Fixing the five test issues (JSON escaping, non-numeric ids, float epsilon, shared-state wiring, zone semantics, credit math):quota + loop green. Fixing the config test base fixture:All green. Quick CLI smoke with a quota-enabled config against the real container `/proc`:Live smoke passes (real /proc, PSI, TZ-correct peak eval). Now docs — example config:Full gate green. Committing and pushing:Pushed. Writing the deliverable REPORT:Done. Delivered on `main` @ `fc518c4` (pushed) + `REPORT-20260829-0500-quota.md` in the PMO inbox.
|
||
|
||
- **Endpoint**: z.ai publishes no usage REST route (probed live + full docs sweep; findings in REPORT) — parser targets the documented bucket model, fake-server tested, `usage_url` flips it on when z.ai ships it; until then the gate runs on credit estimates from the documented formula. LIVE VERIFICATION flagged open.
|
||
- **Decision — Redis container** (not LiteLLM-native): z.ai meters credits, LiteLLM meters $; we own the write path, 9 instances/2 hosts need shared TTL window counters; stdlib RESP2 mini-client keeps the repo dependency-free, fail-soft when down. Runbook + cgroup deploy section in README.
|
||
- **Loop behavior**: resources → block wall (95%) → peak window (01:00–05:00 CST weekdays, verified = Mon–Fri 14:00–18:00 SGT) → soft defer (85%, flash-only) — defer never consumes the task; the 19:00-wall replay test proves defer-and-recover with zero dead turns. Per-class token+credit accounting lands in `loop.jsonl` (`harness quota status|probe|gate`).
|
||
- **Gate**: `./dev.sh check` fully green (11 packages), CLI smoke-tested live in-container against real `/proc`; `[quota]`/`[resources]` off by default so existing configs are untouched (regression-tested).
|
||
A previous attempt already committed `fc518c4`. Assessing completeness before continuing.Gate is clean. Now verifying what the prior attempt left incomplete: report location, push status, CLI subcommand, config surface.Prior attempt's work is committed and pushed (origin/main == `fc518c4`), report delivered, gate clean. Spot-checking the two test surfaces the task demands.Task complete — the prior dispatch's work landed fully; I verified rather than rebuilt:
|
||
|
||
- **Code** `fc518c4` on origin/main: `internal/quota` (parser, TZ-aware schedule w/ embedded tzdata, estimates, RESP2 redis state, resource monitor), `internal/loop` gate integration, `harness quota status|probe|gate`, `[quota]`/`[resources]` config + README tables.
|
||
- **Verified green**: `./dev.sh check` clean across all 11 packages.
|
||
- **Key tests present**: `TestDecideNineteenWall` + `TestLoopQuotaWallDefersAndRecovers` (19:00-wall replay), TZ-edge schedule tables, key-never-logged, redis round-trip vs fake server, usage accounting.
|
||
- **Report** delivered at `~/.coordinate/inbox-pmo/REPORT-20260829-0500-quota.md`: Redis-container decision (stdlib RESP2 client, fail-soft), live endpoint probe (usage routes 404 → estimates + fake server, LIVE VERIFICATION open), config surface, test results.
|