79 lines
4.0 KiB
Markdown
79 lines
4.0 KiB
Markdown
# Efficiency Repair v17
|
|
|
|
## Baseline
|
|
|
|
Source: `efficiency-latest-baseline.json` and `.md`, a fixed 20-cycle sample.
|
|
Phase totals include instrumented retry sleep; older uninstrumented gaps are
|
|
shown separately in the wall-clock timeline. Do not add those gaps twice.
|
|
|
|
| Step | Total seconds | Phase share |
|
|
|---|---:|---:|
|
|
| First response wait | 1114.6 | 42.6% |
|
|
| Provider retry sleep | 930.8 | 35.6% |
|
|
| Initial model decision | 133.6 | 5.1% |
|
|
| Post-activity idle stop | 120.7 | 4.6% |
|
|
| Finalization | 69.1 | 2.6% |
|
|
| Decisions after source reads | 46.2 | 1.8% |
|
|
| Post-turn build | 43.2 | 1.6% |
|
|
| Pre-turn build | 40.7 | 1.6% |
|
|
| Artifact synchronization | 10.7 | 0.4% |
|
|
| Business acceptance | 6.2 | 0.2% |
|
|
| Preparation and packet generation | 8.6 | 0.3% |
|
|
|
|
The detailed report lists the other decision phases, individual commands,
|
|
exit codes, per-cycle waiting and pipeline steps. These are observations of
|
|
different workloads, not a controlled whole-project speedup measurement.
|
|
|
|
## Repairs
|
|
|
|
- Retry sleep previously blocked every scheduler action, including prepared
|
|
native business cases that need no model. The scheduler now permits the
|
|
current ready local pipeline while retaining the model-request backoff.
|
|
Model calls, including periodic optimization, remain behind that backoff.
|
|
- Failed or incomplete local pipelines with unchanged recorded inputs are not
|
|
replayed. Changed source can trigger a new local attempt. Cooled records,
|
|
stale packets and missing business contracts remain protected.
|
|
- Git scope rejections followed only by restricted read-only inspection chains
|
|
now receive machine-verifiable diagnostic dispositions. The parser rejects
|
|
shell substitutions, redirects, fallback operators, mutation commands and
|
|
unknown tails. Original failure entries and observation counts are retained.
|
|
- CAM diagnostic `dd469c1f85e8854d5ded` acquired a diagnostic proof at
|
|
`failure-evidence/diagnostic-20260908T092707979050.json` and no longer needs a
|
|
model review. This is not a new CAD business completion.
|
|
- A current local reconciliation check took 0.085293 seconds and its repeat
|
|
took 0.081562 seconds for 624 records. The raw ledger hash was unchanged;
|
|
31 records were closed and 593 remained pending. These totals include
|
|
earlier work, not 31 new closures from this change.
|
|
|
|
## Checks
|
|
|
|
129 tests passed across scheduler efficiency, business recovery, reconciliation
|
|
efficiency, failure reconciliation and unattended scheduling. Coverage includes
|
|
local work during backoff, no model launch during backoff, unchanged-failure
|
|
suppression, proof invalidation and rejection of unsafe compound shell commands.
|
|
|
|
All four current declared business build targets have a corresponding link rule
|
|
in their existing build tree. Existing source, WASM, build and evidence caches
|
|
are reused; no module was rewritten for this optimization.
|
|
|
|
## Provider Observation
|
|
|
|
The configured provider's model-list request succeeds. A minimal direct
|
|
generation request using the configured model timed out after 20.545 seconds.
|
|
A same-provider `gpt-5.6-sol` text-only probe returned READY, but its subsequent
|
|
read-only tool-call probe did not complete within 60 seconds. This was not
|
|
sufficient evidence for switching the scheduler model; its model configuration
|
|
was retained. The request path remains intermittently unavailable.
|
|
|
|
A text-only request to `/v1/responses` completed in 3.720 seconds, but a subsequent
|
|
CLI tool probe using the production permission mode timed out at 45 seconds.
|
|
A separate read-only-mode probe encountered the local bwrap sandbox restriction.
|
|
The text-only response alone does not establish a working tool execution route.
|
|
With request compression disabled, the same-route tool probe ran `pwd` with
|
|
exit code 0 at 53.470 seconds, but did not emit `turn.completed` within 120 seconds.
|
|
This was insufficient to activate an endpoint or compression override.
|
|
|
|
These repairs remove avoidable model work and unblock eligible local work.
|
|
They do not claim that remote service availability or overall business throughput
|
|
has been restored.
|