pi trial record: pi v0.99.1 behind OmniRoute with the token-save practice (S1, S2, R1 observed, decision trial) - #524

Scope

• What this PR changes: adds the dated record of the pi trial through OmniRoute (blueprints/convergence-practice/pi-omni-trial-20260929/: README, experiment.json, the local kit, sanitized per-run evidence, two gateway read-backs and pool readings) and registers its 28 files and the record in manifests/evidence.json. pi v0.99.1 ran 28 real tasks on sharedgw/gpt-6.1-sol through the 20129 lane (S1 wiring smoke 3, S2 24 = 3 tasks x 4 arms x 2 attempts, R1 one run killed at its second response and resumed); all passed their fixture checks. On these small fixtures the token-save stack used more tokens than plain pi (about 4.9x the uncached input), the gateway lane removed 192 of 413,153 input tokens, and its effect on uncached input is unresolved. Decision stays trial; nothing is adopted, and no verdict row, catalog, pin, client setting or gateway setting changes.
• Base commit: b4056a355580f079560cb612cf03a0a2568c3af7 (rebased onto main after OpenHands recipe lock: PyJWT 2.14.0 relock closes ten advisories (unblocks the required osv-scanner check) #525, the PyJWT relock, and OmniRoute rebuild 2026-09-30: decision record, evidence artifacts, owner-record script, unit templates (records what runs; compression on 20128 not applied) #530, the OmniRoute rebuild record it cites)
• Lane: lane:foundation
• Owned paths touched: blueprints/convergence-practice/pi-omni-trial-20260929/**; manifests/evidence.json (shared hot file: last commit, taken from main and re-registered after the rebase, per docs/lanes.md; the re-registered manifest is byte-identical to the auto-merged one).

SOTA sources

• pi: https://github.com/earendil-works/pi at tag v0.99.1 (d86654abb8862e201933517d6f1fce9f88dd117f): docs/mcp.md, docs/cli.md (tool_search), packages/ai openai-responses.ts:327-337; packages/ai/src/api/anthropic-messages.ts (why the Anthropic OAuth path is excluded).
• RTK: https://github.com/rtk-ai/rtk 0.50.0, rtk init --agent pi. context-mode: https://github.com/mksglu/context-mode 1.0.169, README "Pi Coding Agent" and src/adapters/pi/extension.ts:740-753.
• OmniRoute: https://github.com/diegosouzapw/OmniRoute at 2f42a9ac19d1a247ec9ce5473b790843724b3061 (tree equal to upstream main c1e30b76): open-sse/services/compression/{lossyRequestPolicy.ts:29-71,deriveDefaultPlan.ts,strategySelector.ts:143-173,engineCatalog.ts:200-206,types.ts:421-443}, docs/compression/COMPRESSION_GUIDE.md:233-239, src/app/api/usage/provider-limits/route.ts (POST = manual quota-cache refresh).
• Harbor: https://github.com/harbor-framework/harbor v0.23.0, src/harbor/agents/installed/pi.py lines 88-100, 115-178, 222-235 (why S1/S2 use a local runner and S3 needs a subclass).
• Claude Code legal and compliance page (fetched 2026-09-29): https://code.claude.com/docs/en/legal-and-compliance.

Evidence-class table

Local commands run

$ python3 scripts/validate.py
exit 0  {"components": 69, "hashed_files": 8378, "profiles": 4, "receipts": 174, "status": "passed"}
$ python3 scripts/validate_convergence.py --all-recorded
record[0..25]: valid
$ gitleaks dir blueprints/convergence-practice/pi-omni-trial-20260929
no leaks found
$ (host-string scan of the trial directory: host user name, /home/, /tmp/, session id) 0 matches

Independent reviews, two families, one repair round each: (1) an Opus evidence-reviewer (read-only): no blocker, 9 should-fix, all applied (the "upstream-default" label corrected to "upstream's documented safe lane plus ccr", cache-warmth confound disclosed, unarchived figures removed or labelled, the 14:12Z run's role and compression note corrected, read-back extended to the full key set). (2) A GPT-6 astra review at max through OmniRoute (20129 to 20128, compression off, read-only sandbox): no blocker, every recomputed number confirmed (S2 table, request-2 shares, paired differences, R1 sums, all 29 usage objects, 47 hashes); about 11 should-fix and 10 minor findings, applied in the second commit: analysis scripts (summarizer merged the gateway-off control into stack, no stage selection, null usage read as zero, silent row truncation, R1 export overwrite, run exit statuses masked), the mutant self-test dropped verdicts from the probe tally ("of 98" is now "of 100"), wording scoped to what n=2 fixtures support, and two undisclosed confounds now disclosed (stack and stack-ext shared Serena/jCodeMunch state directories; the header-less arms passed through the gateway's lite whitespace fold). Not changed, by design: the runners and templates that produced the evidence (their hashes must match the runs); their findings (inherited PI_CODING_AGENT_DIR, PID-by-PID kill) are disclosed in the record. Residuals: no live re-run of R1 with the hardened kill; provenance of the executed run_s2.sh predecessor.

Decision record

The record's own decision is trial (experiment.json); its next_decision_changing_test is a large-output fixture (S2b) with the same four arms.

Host evidence

No files under evidence/hosts/ change.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no workflow changes).
• New/changed workflows declare top-level permissions: contents: read (no workflow changes).
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted, moved or overwritten).

🤖 Generated with Claude Code