devlog: round-3 Cursor live-probe evidence and lock (docs-only)

devlog/_plan/260822_senpi_cursor_transfer/200_round3_probe_plan.md

@@ -0,0 +1,58 @@
+# 200 — Round-3 live-probe plan
+
+Round 2 (100-190) landed T04/T07/shutdown. The 190 lock deferred five rows
+for lack of live evidence (T06 maxMode, T02 rotation, #2305 external
+continuation, client version watch, T09 cacheRead clamp); T08/T10 stay
+deferred for non-live reasons and #2334 stays NEEDS_HUMAN by design. A live
+Cursor account now exists on the probe host (macmini, ocx preview), so this
+cycle buys the evidence.
+
+## Probes
+
+1. **P-1 maxMode (T06).** Dump GetUsableModels with full ModelDetails —
+   which models report maxMode=true, and what contextTokenLimit pairs with
+   it. Compare a Run with RequestedModel.maxMode=true vs false on one
+   maxMode-capable model: does the server accept it, and does the reported
+   context window / usage change? Wire flag only lands if this shows a real
+   user-visible gain.
+2. **P-2 rotation (T02 / #1527 suspect).** Drive a conversation toward the
+   bare 0-token resource_exhausted shape (large-context turns on a pinned
+   conversationId). If the server pins the rejection to the conversationId
+   (fresh id succeeds with identical payload), T02 rotation is justified;
+   implement bounded rotation + checkpoint rekey. If not reproducible within
+   quota bounds, record and keep deferred.
+3. **P-3 issue #2305.** Reproduce the client-tool continuation returning
+   tool-call-like assistant text to Pi: drive a client-tool turn through the
+   external continuation path and capture what text frames come back.
+   Root-cause lives in rootPromptMessages / userMessageAction continuation
+   (the a69d291fb fix covered native Auto; #2305 is the external path).
+4. **P-4 client version.** GetUsableModels + one Run with the current pinned
+   cli-2026.07.08-0c04a8a vs a newer senpi-observed string
+   (cli-2026.07.23-e383d2b): any catalog or behavior delta? Bump only if
+   probe shows the new string is accepted and changes nothing adverse.
+5. **P-5 billed usage / cacheRead (T09).** Capture the billed turnEnded
+   usage int64s (inputTokens / outputTokens / cacheRead*) from the SAME live
+   Runs P-1 and P-4 already make (no extra quota): decode and record whether
+   cacheRead exceeds 3x input the way senpi's clamp assumes, and whether our
+   protobuf-events usage mapping already reports these fields sanely. Verdict
+   IMPLEMENT (clamp justified) / NOOP (values sane, clamp unnecessary) /
+   BLOCKED (fields absent on this plan tier).
+
+## Probe hygiene (binding, extends doc 100 boundary)
+
+- All transcripts REDACTED before entering devlog: no bearer tokens, no
+  account ids, no email, no checksum headers. Raw dumps stay in .tmp/ on the
+  probe host and are deleted after the docs lock.
+- Quota respect: P-2 large-context attempts are capped (<= 5 runs); if the
+  account rate-limits, stop and record BLOCKED for that probe.
+- No Safe Storage access, no client patching, no endpoints beyond what the
+  adapter already ships (Run, GetUsableModels, RunSSE fallback).
+
+## Outputs
+
+- 210_maxmode.md, 220_rotation.md, 230_issue2305.md, 240_client_version.md —
+  each with verdict IMPLEMENT / NOOP / BLOCKED / NEEDS_HUMAN and, for
+  IMPLEMENT, diff-level shape.
+- 250_billed_usage.md — P-5 verdict for T09 (same contract).
+- 290_round3_lock.md — ranked implementation order + updated senpi
+  superiority verdict.

devlog/_plan/260822_senpi_cursor_transfer/210_maxmode.md

@@ -0,0 +1,28 @@
+# 210 — P-1 maxMode probe (T06)
+
+## Evidence (macmini live, 2026-08-22, redacted)
+
+- GetUsableModels decoded: 204 entries; ModelDetails keys =
+  modelId, displayModelId, displayName, displayNameShort, aliases, maxMode.
+- maxMode=true on exactly 28 ids — ALL of them opus "-fast" variants
+  (claude-opus-5-*-fast, claude-opus-4-8-*-fast, claude-opus-4-7-*-fast).
+  No contextTokenLimit field is present in this response shape.
+- Run A/B on claude-opus-4-7-low-fast, tiny prompt:
+  - RequestedModel.maxMode=false -> bare Connect resource_exhausted.
+  - RequestedModel.maxMode=true  -> same bare resource_exhausted.
+  The server ACCEPTED the flag both ways (no invalid_argument); the model is
+  plan-gated for this account regardless.
+
+## Verdict: BLOCKED (plan tier)
+
+maxMode only decorates -fast (paid burst) variants, and this account cannot
+run them at all, so no user-visible gain is provable here. Wire flag stays
+hardcoded false. Re-probe requires an account with -fast entitlement
+(NEEDS_HUMAN to provision).
+
+## Side finding (feeds 260)
+
+A TINY prompt on a plan-gated model returns the same bare 0-token
+resource_exhausted shape that #2320 (T01) now classifies as CONTEXT OVERFLOW.
+Live proof that bare RE != always overflow: entitlement rejections share the
+shape. See 260_re_classification_refinement.md.

devlog/_plan/260822_senpi_cursor_transfer/220_rotation.md

@@ -0,0 +1,15 @@
+# 220 — P-2 rotation probe (T02 / #1527 suspect)
+
+## Evidence
+
+4 consecutive ~101K-token turns on one pinned conversationId
+(composer-2.5-fast) all completed (OK1..OK4, usage.totalTokens ~101,111-147).
+No 0-token resource_exhausted, no conversation poisoning within the capped
+attempt budget (probe cap <= 5 runs, quota hygiene doc 200).
+
+## Verdict: NOT REPRODUCED — T02 stays deferred
+
+The senpi #998 pathology (server pinning a rejection to a conversationId) did
+not manifest at this size on this plan. #1527 remains open without a local
+reproduction; rotation-with-persistence stays deferred until a live
+reproduction exists. No implementation this round.

devlog/_plan/260822_senpi_cursor_transfer/230_issue2305.md

@@ -0,0 +1,27 @@
+# 230 — P-3 issue #2305: display-alias leak in assistant text
+
+## Root cause (code-grounded, ox-alpha lane + main-agent verification)
+
+OpenCodex has NO text-mode tool-call parser; assistant text passes through
+verbatim: protobuf-events.ts textDelta (~:1245) -> message-mapper.ts ->
+bridge.ts text_delta -> chat-completions client. Pi parses
+"[TOOL_CALL]name[ARGS]{...}" text itself, so when a Cursor model emits the
+textual pseudo-frame with the DISPLAY name (mcp_opencodex-responses_grep),
+Pi sees an undeclared tool and the turn dies. Real tool-call FRAMES are
+already normalized via mcpWireNameFromArgs -> normalizeCursorWireName
+(protobuf-events.ts:278-281); text deltas bypass that.
+
+## Verdict: IMPLEMENT
+
+## Diff shape
+
+- protobuf-events.ts textDelta case: scrub via marker-scoped regex
+  \[TOOL_CALL\](mcp_opencodex-responses_[^\[\]]+)\[ARGS\] ->
+  normalizeCursorWireName inside markers only. Prose mentions stay untouched;
+  scope-guarded to the exact OCX_RESPONSES_TOOL_PROVIDER prefix.
+- Streaming caveat: a marker can straddle two deltas. Start WITHOUT tail
+  buffering; add only if live traces show split markers (recorded risk).
+- Tests: tests/cursor-protobuf-events.test.ts — marker normalized, prose
+  untouched, real frames unaffected.
+- Precedent: a69d291fb (request-side [Tool Result] envelope strip) — same
+  failure family, response-side analogue.

devlog/_plan/260822_senpi_cursor_transfer/240_client_version.md

@@ -0,0 +1,14 @@
+# 240 — P-4 client version probe
+
+## Evidence
+
+GetUsableModels accepted all three version strings with byte-identical
+catalogs (204 entries): cli-2026.07.08-0c04a8a (ours), cli-2026.07.23-e383d2b
+(senpi), cli-2026.02.13-41ac335 (our discovery pin). Live Run on the 07.08
+pin works (P-5 turns completed).
+
+## Verdict: NOOP (no forced bump)
+
+No behavioral delta proven. Optional freshness bump to 07.23 is safe by this
+probe but buys nothing measurable; keep the pin, keep the drift watch from
+190 (api2direct reports).

devlog/_plan/260822_senpi_cursor_transfer/250_billed_usage.md

@@ -0,0 +1,13 @@
+# 250 — P-5 billed usage / cacheRead (T09)
+
+## Evidence
+
+Two live proxy turns (composer-2.5-fast) report sane Responses usage:
+input_tokens 11085/11162, output 11/10, cached_tokens 0, no inflation, no
+cacheRead > 3x input pathology. Transport-level runs report estimated usage
+consistently (~101K totals on the big turns, matching payload size).
+
+## Verdict: NOOP for the clamp
+
+No evidence of senpi's billed-int64 pathology on this plan tier. T09 clamp
+stays unimplemented; revisit only if live usage reports regress.

devlog/_plan/260822_senpi_cursor_transfer/260_re_classification_refinement.md

@@ -0,0 +1,31 @@
+# 260 — bare resource_exhausted refinement (T01 follow-up, live-evidenced)
+
+## Problem
+
+#2320 classifies a bare 0-token resource_exhausted (no quota cue, no size
+phrase) as CONTEXT OVERFLOW -> 400-class so Codex compacts. Live probe 210
+found a counterexample: a ~20-token prompt on a plan-gated model
+(claude-opus-4-7-low-fast) returns the SAME bare shape. Misclassifying that
+as overflow makes Codex compact a 20-token turn — wrong remedy, confusing UX,
+and the retry can never succeed.
+
+## Design
+
+Classification needs a size prior: only classify bare RE as overflow when the
+REQUEST was plausibly large relative to the model's context window; small
+requests keep the 429-class quota/entitlement mapping. The adapter already
+computes an input-token estimate (prepareCursorRunRequest
+estimateInputTokens; estimateTokens lib). Shape:
+
+- cursor-errors.ts: classifyCursorError gains an optional context
+  { estimatedInputTokens?, contextWindow? }.
+- live-transport/adapter passes the estimate it already has for the turn.
+- Rule: bare RE + estimate >= OVERFLOW_MIN_FRACTION (0.5) * contextWindow ->
+  overflow (current behavior); otherwise -> existing rate-limit mapping.
+  Unknown estimate/window -> keep current overflow mapping (fail toward
+  compaction, today's behavior) so the refinement only ever REDUCES
+  false overflows it can prove.
+- Tests: tests/cursor-errors.test.ts — tiny-estimate bare RE -> 429 class;
+  large-estimate -> overflow; no-estimate -> overflow (unchanged).
+
+## Verdict: IMPLEMENT (beyond-senpi refinement; senpi T01 shares this bug)

devlog/_plan/260822_senpi_cursor_transfer/290_round3_lock.md

@@ -0,0 +1,31 @@
+# 290 — Round-3 lock
+
+Probes executed on macmini (live account, redacted transcripts in 210-250;
+raw dumps deleted from probe host after lock per 200 hygiene).
+
+## Implementation order (this loop)
+
+1. **230 — #2305 text-marker normalization** (IMPLEMENT; clear defect, open
+   issue, code-grounded fix point).
+2. **260 — bare-RE size prior** (IMPLEMENT; live-evidenced false-overflow
+   class; a refinement senpi's own T01 lacks).
+
+## Closed by probe (no code)
+
+- 210 maxMode: BLOCKED (plan tier) — flag accepted but -fast entitlement
+  absent; NEEDS_HUMAN to provision a -fast-capable account for re-probe.
+- 220 rotation: NOT REPRODUCED at 4x101K; T02 stays deferred.
+- 240 client version: NOOP — three version strings byte-identical catalogs.
+- 250 billed usage: NOOP — no cacheRead pathology on this plan.
+
+## Updated senpi verdict
+
+With 230+260 landed, remaining senpi-ahead rows shrink to: rotation
+persistence (unreproducible here), maxMode (plan-gated for both projects
+without entitlement), agent-loop-level stop/exec ownership (out of adapter
+scope; #2305's actual defect is ours to fix and is fixed). OpenCodex keeps
+its unique-side advantages (interactionQuery, HTTP/1 fallback, SelectedImage
+vision, bounded memory, T04 watchdog with senpi-matching thresholds, typed
+exec errors, EOF fail-closed tests). Verdict: at parity or ahead on every
+row that is provable on this plan tier; the two rows senpi still leads
+require entitlement or a reproduction neither project can show today.