GeezerVison

Idle compaction now remembers the model from each session's latest main usage event and selects a 30-minute fallback for GPT-5.6 and later on official OpenAI API and ChatGPT providers. Cache-lifetime learning now uses the last sufficiently sampled group with reliable hits rather than stopping at the first failing group. The change also adds regression tests, updates the cache documentation, and records the beta and stable releases.

Changes

• Files reviewed: 7 · 76 changed lines · estimated review effort 2/5, about 15 minutes
• Profile: normal · at most 8 inline comments
• Verification: the review pass raised no candidate findings, so no separate pass was needed
• Cost: 350,653 tokens in (306,816 from the provider's cache) · 2,059 out
• Readers: 4 files went to readers instead of the review pass whole: 4 code files by gpt-6-luna (the review pass gets their summaries and the hunks of their risk spots; 1 call, up to 3 at once, 12,521 tokens counted). Their summaries are hints, not findings. 3 code files no reader could read went to the review pass whole (the reader gave no usable answer for this file).
• Stacked on #144: only this pull request's own commits since #144's head 479cd5a.

Performance (perf-attrib, base 479cd5a against head fd777e2)

perf-attrib: sube: tui.after.fp_mb +7.9MB; total.after.fp_mb +7.9MB; tui.ready.fp_mb +3.3MB. Causes: tui.after.heap.type.Function +3; tui.rest.heap.type.Function +2; tui.rest.heap.type.StructureChain +2; tui.after.heap.type.JSLexicalEnvironment +1; tui.rest.ctor.native:Blob.bytes +725B; tui.rest.ctor.object:ModuleRecord.bytes +308B; tui.rest.ctor.code:NativeExecutable.bytes +80B

base 479cd5a694c1a170b8759fedbdc3edc2627969bd (479cd5a6) vs cand fd777e2c157f0d059102ad33101c96fce2b2bc27 (fd777e2c), median of 2 faux-engine session(s) per side; flagged = beyond the noise measured on 18 base-vs-base runs (x1.5).

JS object types that rose (count):

• tui.after.heap.type.Function: 44524 -> 44527 (+3)
• tui.rest.heap.type.Function: 44521 -> 44523 (+2)
• tui.rest.heap.type.StructureChain: 90 -> 92 (+2)
• tui.after.heap.type.JSLexicalEnvironment: 24574 -> 24575 (+1)

JS heap bytes by constructor that rose:

• tui.rest.ctor.native:Blob.bytes: 1742388 -> 1743113 (+725)
• tui.rest.ctor.object:ModuleRecord.bytes: 5504250 -> 5504558 (+308)
• tui.rest.ctor.code:NativeExecutable.bytes: 89120 -> 89200 (+80)

Modules loaded at start: 230 -> 230, 4632729 -> 4633037 bytes of chunks; chunk sizes only in cand: [1008970, 110586], only in base: [1008942, 110306] (chunk names are hashes: the list says which changed in size, not by what name)

Reference (skill, 2026-11-03, real model): footprint 64 ready / 66 after / peak 80 MB, 2.6 KB per session. This run (faux engine, whole process tree): 58.1 / 65.3 / peak 73.6 MB, 2058 B per session.

The Telegram bridge now identifies session buttons by session ID, preserves shared subscriptions while another chat still uses them, and clears stale questions when leaving a session. Watching a running session first reads its history and then replays events from the turn's start, with per-chat sequence tracking intended to suppress duplicates. Separately, the native engine applies existing bare-task limits to file mentions and to system-prompt and context extension handlers; idle-compaction and release changes are outside this incremental diff.

Changes

Sequence diagram

sequenceDiagram
    participant Chat as Telegram chat
    participant Bridge as Channel bridge
    participant Client as Shared daemon client
    participant Daemon
    Chat->>Bridge: Pair or select session
    Bridge->>Daemon: Read recap
    Bridge->>Bridge: Assign target session and reset seen
    Bridge->>Client: watch(sessionId)
    Client->>Daemon: session.history(limit=1)
    Daemon-->>Client: Latest turn
    Client->>Daemon: Subscribe before running turn, or after latest window
    Daemon-->>Client: Replay and live events
    Client-->>Bridge: onEvent
    Bridge-->>Chat: Prompt, progress and approval messages

• Files reviewed: 8 · 208 changed lines · estimated review effort 3/5, about 30 minutes
• Profile: normal · at most 8 inline comments
• Verification: every finding was checked again by a separate pass (2 candidates judged)
• Cost: 695,967 tokens in (600,576 from the provider's cache) · 4,740 out
• Readers: 6 files went to readers instead of the review pass whole: 6 code files by gpt-6-luna (the review pass gets their summaries and the hunks of their risk spots; 1 call, up to 3 at once, 16,445 tokens counted). Their summaries are hints, not findings. 2 code files no reader could read went to the review pass whole (the reader gave no usable answer for this file; the reader flagged more risk spots than a summary carries, so the file goes whole).

Performance (perf-attrib, base 479cd5a against head 80f6e50)

perf-attrib: igual: nothing above the measured noise

base 479cd5a694c1a170b8759fedbdc3edc2627969bd (479cd5a6) vs cand 80f6e501efd1be3ffce29580286258750fd71cba (80f6e501), median of 2 faux-engine session(s) per side; flagged = beyond the noise measured on 18 base-vs-base runs (x1.5).

JS object types that rose (count):

• tui.after.heap.type.Function: 44526 -> 44527 (+1)
• tui.rest.heap.type.Function: 44522 -> 44523 (+1)

JS heap bytes by constructor that rose:

• tui.rest.ctor.native:Blob.bytes: 1742388 -> 1743266 (+878)
• tui.rest.ctor.object:ModuleRecord.bytes: 5504250 -> 5504809 (+559)

Modules loaded at start: 230 -> 230, 4632729 -> 4633288 bytes of chunks; chunk sizes only in cand: [1009142, 251528, 110586], only in base: [1008942, 251449, 110306] (chunk names are hashes: the list says which changed in size, not by what name)

Reference (skill, 2026-11-03, real model): footprint 64 ready / 66 after / peak 80 MB, 2.6 KB per session. This run (faux engine, whole process tree): 54.7 / 56.9 / peak 73.3 MB, 2040 B per session.

The supplied incremental diff changes Telegram session attachment, not idle compaction or release metadata. A new join phase buffers target-session events while the watch starts; a successful attachment resets session-local state and delivers buffered events in sequence order. Failed switches retain the previous session's mirrored turn and pending questions. The same attachment path is used for session creation, restoration, and explicit switching.

Changes

Sequence diagram

sequenceDiagram
    participant Chat as Telegram chat
    participant Bridge as Channel bridge
    participant Host as Daemon host
    Chat->>Bridge: Select session
    Bridge->>Host: watch(target)
    Host-->>Bridge: Replay and live events
    Note over Bridge: Buffer target events; retain previous session
    Host-->>Bridge: Watch result
    alt Watch succeeds
        Bridge->>Bridge: Enter target and deliver buffered events by seq
        Bridge-->>Chat: Mirrored output and requests
    else Watch fails
        Bridge-->>Chat: Report failure; preserve previous session
    end

• Files reviewed: 2 · 100 changed lines · estimated review effort 2/5, about 15 minutes
• Profile: normal · at most 8 inline comments
• Verification: every finding was checked again by a separate pass (1 candidate judged)
• Cost: 254,191 tokens in (199,168 from the provider's cache) · 2,861 out
• Readers: 1 file went to readers instead of the review pass whole: 1 code file by gpt-6-luna (the review pass gets their summaries and the hunks of their risk spots; 1 call, up to 3 at once, 9,646 tokens counted). Their summaries are hints, not findings. 1 code file no reader could read went to the review pass whole (the reader flagged more risk spots than a summary carries, so the file goes whole).

Performance (perf-attrib, base 479cd5a against head 7c7dfba)

perf-attrib: igual: nothing above the measured noise

base 479cd5a694c1a170b8759fedbdc3edc2627969bd (479cd5a6, cached build) vs cand 7c7dfbaada2d109b7c06fbcdfbeb7f80c68225ef (7c7dfbaa), median of 2 faux-engine session(s) per side; flagged = beyond the noise measured on 18 base-vs-base runs (x1.5).

JS object types that rose (count):

• tui.after.heap.type.Structure: 16373 -> 16505 (+132)
• tui.after.heap.type.Function: 44526 -> 44527 (+1)
• tui.rest.heap.type.Function: 44522 -> 44523 (+1)

JS heap bytes by constructor that rose:

• tui.rest.ctor.native:Blob.bytes: 1742388 -> 1743266 (+878)
• tui.rest.ctor.object:ModuleRecord.bytes: 5504250 -> 5504809 (+559)

Modules loaded at start: 230 -> 230, 4632729 -> 4633288 bytes of chunks; chunk sizes only in cand: [1009142, 251528, 110586], only in base: [1008942, 251449, 110306] (chunk names are hashes: the list says which changed in size, not by what name)

Reference (skill, 2026-11-03, real model): footprint 64 ready / 66 after / peak 80 MB, 2.6 KB per session. This run (faux engine, whole process tree): 55.1 / 57.7 / peak 73.7 MB, 2041 B per session.

The bridge now passes the chat ID into its session-joining helper and checks, after the asynchronous watch completes, that the runtime still belongs to a paired chat. If pairing was revoked, it discards buffered events, releases a successful watch, and returns a distinct cancellation result so callers avoid creating a replacement session or sending a misleading watch-failure message. The incremental diff changes only the Telegram bridge and its test coverage; it does not change idle compaction or release metadata.

Changes

Sequence diagram

sequenceDiagram
    participant Chat as Telegram chat
    participant Bridge as Channel bridge
    participant Host as Session host
    participant Control as Terminal control
    Chat->>Bridge: Select session
    Bridge->>Host: watch(sessionId)
    Host-->>Bridge: Session events buffered during startup
    Control->>Bridge: Remove pairing and forget runtime
    Host-->>Bridge: Watch succeeds
    Bridge->>Bridge: Detect revoked pairing or stale runtime
    Bridge->>Host: unwatch(sessionId) unless another chat follows it
    Bridge-->>Chat: Answer callback: Not switched

• Files reviewed: 2 · 33 changed lines · estimated review effort 1/5, about 5 minutes
• Profile: normal · at most 8 inline comments
• Verification: the review pass raised no candidate findings, so no separate pass was needed
• Cost: 133,100 tokens in (104,704 from the provider's cache) · 1,338 out
• Readers: 2 code files no reader could read went to the review pass whole (the diff is small: one reader call costs more than the readers' share of this review's cap, so the review pass reads it).

Performance (perf-attrib, base 479cd5a against head 1aed947)

perf-attrib: nota: no total rose; inside, tui.ready.heap.objects +120

base 479cd5a694c1a170b8759fedbdc3edc2627969bd (479cd5a6, cached build) vs cand 1aed947efa71e3694ca6a5175cd0cb796b421fc6 (1aed947e), median of 2 faux-engine session(s) per side; flagged = beyond the noise measured on 18 base-vs-base runs (x1.5).

JS object types that rose (count):

• tui.after.heap.type.Structure: 16469 -> 16508 (+39)
• tui.ready.heap.type.UnlinkedFunctionExecutable: 14131 -> 14155 (+24)
• tui.ready.heap.type.FunctionExecutable: 14145 -> 14168 (+23)
• tui.ready.heap.type.SymbolTable: 3406 -> 3416 (+10)

JS heap bytes by constructor that rose:

• tui.rest.ctor.native:Blob.bytes: 1742388 -> 1743266 (+878)
• tui.rest.ctor.object:ModuleRecord.bytes: 5504250 -> 5504809 (+559)

Modules loaded at start: 230 -> 230, 4632729 -> 4633288 bytes of chunks; chunk sizes only in cand: [1009142, 251528, 110586], only in base: [1008942, 251449, 110306] (chunk names are hashes: the list says which changed in size, not by what name)

Reference (skill, 2026-11-03, real model): footprint 64 ready / 66 after / peak 80 MB, 2.6 KB per session. This run (faux engine, whole process tree): 54.6 / 57.3 / peak 74.1 MB, 2039 B per session.

• Files reviewed: 0 · 0 changed lines · estimated review effort 1/5, about 5 minutes
• Profile: normal · at most 8 inline comments
• Verification: the review pass raised no candidate findings, so no separate pass was needed
• Cost: 0 tokens in · 0 out
• No code graph index: reviewed from the diff alone.