Skip to content

History / Harnesses

Revisions

  • harnesses: recorded the per-platform lease primitive and its measurement - The lease is `flock` on macOS and an open-file-description lock on Linux since omp 19.0.1; Linux keeps the two lock spaces apart, so a `flock` probe reports a live lease as free (measured in a container: `flock` free, `F_OFD_GETLK` `F_WRLCK`). - Naming the primitive in the panel page too, since it is the file the stats plugin walks.

    @rajebdev rajebdev committed Nov 20, 2026
  • docs(harnesses): document OmO over the pi-acp bridge - Added the `omo` transport row: no native ACP server, driven through `pi-acp` with `PI_ACP_PI_COMMAND=omo`. - Recorded the measured surface on 5.1.9 (69 models, 3 thought levels, list/delete) and the fork/transcript facts.

    @rajebdev rajebdev committed Nov 19, 2026
  • docs(harnesses): record a server-spawned ACP child's transcript - Documented the empty-transcript bug measured on a real antigravity worker: `chat_sessions.messages='[]'`, `origin=''`, a parent relation, and the answer only in `session_answers`. - Recorded the two halves: the `AcpTranscriptRecorder` writing `acp_transcripts` and the loader/restore gate reading the local transcript instead of `origin`.

    @rajebdev rajebdev committed Nov 18, 2026
  • docs(harnesses): document the deleted-session tombstone

    @rajebdev rajebdev committed Nov 17, 2026
  • docs(harnesses): list the stores the transcript guard reads

    @rajebdev rajebdev committed Nov 17, 2026
  • docs(harnesses): note that a non-empty dsh file is not a transcript

    @rajebdev rajebdev committed Nov 17, 2026
  • docs(harnesses): document the transcript guard for imports and the sidebar

    @rajebdev rajebdev committed Nov 17, 2026
  • docs: record agent-side session deletion and the batch split

    @rajebdev rajebdev committed Nov 17, 2026
  • docs(harnesses): recorded terminal session import and restore - Added the session/list import rules: no cwd (codex tears its app-server down), both-id dedupe, origin marker, and the default-on auto pass with its per-harness switch, boot/sidebar/15-minute cadence, and skip list. - Documented replayFrom restore and the Codex rollout fallback for a thread another client still owns (thread_active_writer). - Recorded the v2 mcpServers shape (tagged entries, absolute stdio command) and the switch-label theme rule.

    @rajebdev rajebdev committed Nov 17, 2026
  • docs(harnesses): recorded the @harness:<id>: model completion

    @rajebdev rajebdev committed Nov 16, 2026
  • docs(harnesses): document the Hermes ACP harness Adds the transport row and the fork row for `hermes`, with the facts a future reader would otherwise re-break: the `acp` extra is a separate sync step on a stock install, the adapter reserves stdout for JSON-RPC, `fork_session` takes no anchor so the fork is whole-chat, and the provider table in `~/.hermes/config.yaml` is the agent's own writer.

    @rajebdev rajebdev committed Nov 14, 2026
  • docs(harnesses): record the per-harness access tier and its measured leak The composer's access chip stored one global value, so `Full bypass` picked on omp followed the operator to `claude` and auto-answered that agent's approval requests. The copy was omp's too, promising `tools.restrictedFolders` behaviour an ACP agent does not have. Document the rule and the measurement: the pick is per harness, omp's entry is mirrored into the server's spawn-time default, every other harness starts at `always-ask`, and the copy follows the transport.

    @rajebdev rajebdev committed Nov 13, 2026
  • docs(harnesses): record the launchd PATH measurement behind "not installed" The boot probe's availability check searched the inherited PATH, and a login entry gives it `/usr/bin:/bin:/usr/sbin:/sbin` — measured, the plist carries no `EnvironmentVariables`. omp survived through its own fallback list, so the first boot after login probed only `omp-acp` and every `npx`-launched harness read "not installed"; a terminal start probes the whole roster. Documents the repair (`withToolchainPath`) that spawns already had, the one lookup that now owns resolution, and the end-to-end verification under the launchd PATH.

    @rajebdev rajebdev committed Nov 13, 2026
  • docs(harnesses): record the jcode and grok ACP measurements Adds both agents to the transport and fork tables with what was measured rather than what the docs claim: - jcode 0.99.1 speaks ACP through `jcode acp` (loadSession plus sessionCapabilities resume/close, no fork) and its providers live in `~/.jcode/config.toml`. - grok 1.0.77 ships `grok agent stdio`: loadSession, list/resume/close, mcpCapabilities stdio+http+sse, and session/new returns standard configOptions plus a `models` block. The BYOK note is the useful part: a `[model.<name>]` table in `~/.grok/config.toml` with its own base_url/api_key runs without an xAI login, whereas GROK_MODELS_BASE_URL + XAI_API_KEY alone still answered "Not signed in". End-to-end on the kenari endpoint, grok answered `hello` on deepseek-v4-1-flash (11 787/3 tokens) and jcode answered `hello` (2 940/3).

    Wijanarko Putra Rajeb committed Nov 12, 2026
  • docs: document adding a harness - A harness is a `custom_harnesses` row plus a descriptor in the shared overlay; the ACP transport was already generic. - `HarnessId` is open, the complete tables are keyed by `BuiltinHarnessId`, and every fact has a fallback (MCP → acp, providers → none, updates → none, fork → none, transcript → false). - A removed harness keeps its id on old sessions, so resume refuses by name instead of rewriting the chat to `omp-rpc`.

    @rajebdev rajebdev committed Nov 11, 2026
  • docs(fork): record the three-tier capability model and the Fork chat action forkCapability is message/tip/none; the row button follows canForkAt and the session menu's Fork chat covers every harness that advertises session/fork, probe-authoritative. The page also names the upstream gaps (opencode, minimax, kimi) so a future reader does not mistake them for a Kabmo bug.

    @rajebdev rajebdev committed Nov 11, 2026
  • docs(fork): record the per-harness fork point and the button's gating Standard ACP has no fork point; claude-agent-acp and codex-acp read the AIR _meta.jetbrains.air.fork extension, omp is cut on disk, and the rest fork the whole session. The page now carries the measured table (advertised fork, fork point, transcript store, Kabmo path) and the rule that canForkAt gates the row button, so a future reader does not re-add a control the send path refuses.

    @rajebdev rajebdev committed Nov 11, 2026
  • docs(harnesses): record the ACP prewarm and the pending chat's mode ladder The ACP pool warms a PROCESS and stops at `initialize`: measured, `omp-acp` writes its session JSONL at `session/new` and the sidebar scans that directory, so a warm child that nobody claimed showed up as an empty chat row. The page carries the split that decision rests on (spawn ~1 ms, `initialize` ~100 ms, `session/new` ~80–120 ms), the smoke numbers (cold 208 ms, warm claim 89 ms, zero session files), and the reason a warm ACP process can serve any pick while omp's cannot. Also records the pending chat's mode source: the harness probe's ladder, with an empty `currentValue` when neither the operator nor the probe named a mode.

    @rajebdev rajebdev committed Nov 11, 2026
  • docs(harnesses): record the composer's ACP mode selector The Session Modes row now names the whole path: the /modes read of the live selector, the dropdown that replaces the Plan toggle for agents with more than one mode, the optimistic set_mode write and its rollback, and the frame-driven re-read. It also states why omp-acp keeps the Plan toggle.

    @rajebdev rajebdev committed Nov 10, 2026
  • docs(harnesses): record ACP v1/v2 support and the client surface The harness page now names what Kabmo advertises on a v1 connection, how the hybrid initialize keeps a v1 peer working, and the v2 adapter rules: capability normalization, setup configId/commands, the prompt lifecycle moving to state_update, permission subjects, and the one-upsert tool model. It also lists the v2 surfaces deliberately left unadvertised (elicitation, auth login, agent-owned terminals, compaction) and the background-update limitation of the current timeline.

    @rajebdev rajebdev committed Nov 10, 2026
  • docs(harnesses): document the ACP background subagent probe Record the measured antigravity behaviour (agy 1.5.3): invoke_subagent ends the turn while the subagent keeps working, and completion updates are held until the next request (google-antigravity/antigravity-cli#1461). Kabmo now probes session/load, swallows the replay, and emits the stable text past the holding answer as a follow-up run. Also records that empty session/prompt is refused and session/resume does not subscribe.

    @rajebdev rajebdev committed Nov 10, 2026
  • docs(harnesses): document the background update worker and its log The update is no longer run inside the POST: a detached worker owns the steps, writes `run/harness-update.json` (spec + record + flock lock), answers 202/409 from the record, and applies the post-install side effects once on the next status read. The page records the three measured failure modes of the old shape (request-lifetime record, the five-minute timeout killing a 147 MB install mid-write, no single flight), the `--allow-scripts=<package>` postinstall fix that explains the "claude native binary not installed" stub, the 30-minute per-step kill bound, the dead-reporter settle rule, and where the About card's "View log" panel reads from.

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harness): record ACP rewind and New Chat harness inheritance The Undo/Retry rule is a capability, not a transport: `omp-rpc` and `omp-acp` share the session JSONL, and the endpoint tears a live ACP child down before truncating. The other ACP agents are refused before their child is touched. New Chat captures the source chat's harness on open and carries it into the first spawn, so an answer from an ACP agent does not start an omp-rpc chat; the omp prewarm is skipped because its pool only holds RPC children.

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harnesses): record the ACP footer model and silent-turn rules The footer kept the first model because ACP messages carried no model and an idle chat trusted the last run's sidebar row; a silent turn only reached console.error. Record both root causes, the set_config_option {} fallback, the live-row rename, and the fake-agent check so the next change does not reintroduce either.

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harness): record zcode's provider_config rules and the measured request map ZCode 3.12+ personal providers live in ~/.zcode/v2/provider_config.json, not the app's legacy config.json. Document the rule graph, the model-rule requirement, the empty reasoning map measured against a gateway that rejected enable_thinking with 400, and the runtime requirement (desktop app or zcode-app-cli).

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harness): record why CLI-less adapter rows and pinned npx updates were wrong Names the measurements: zcode/claude rendered a dash over a known adapter number; a global adapter 0.101.0 could not affect a launch pinned to 0.100.0; dsh's ACP surface 0.0.1 is not its CLI package version 0.2.2-rc.1.

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harnesses): record the Claude gateway alias and system-role measurements The env box names one `ANTHROPIC_DEFAULT_<FAMILY>_MODEL` per picker alias. Measured with the SDK binary: an unpinned `--model sonnet` still sends the vendor id, and a gateway that lists its own ids answers `400 model not found`. Also recorded: Claude Code travels the Environment block as a `role: "system"` entry inside `messages`. A gateway that forwards that role to an upstream accepting only user/assistant fails EVERY turn with `400 the model's provider rejected the request` — model-dependent, not an env-form defect (kenari: qwen3-8-27b:free refused, agnes-2-5-flash:free answered).

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harnesses): record that ACP messageId is optional Antigravity 1.5.3 sends agent_message_chunk with no messageId, so the old mapper minted a fresh client id per chunk and rendered one turn as N duplicate rows. Document the measurement, the per-turn synthesized identity, and the tests that pin it.

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harnesses): record registry refresh and the models topic as invalidation Two measured rules for the provider-write path: the pooled `omp` child loads `models.yml` at spawn (so Kabmo's cache alone answered the old list), and a harness catalog that ignored the `models` topic kept pre-write probes until a reload. Also the cost of `realtimeClient.subscribe` replaying the current topic value, and why the second catalog read on mount is accepted.

    @rajebdev rajebdev committed Nov 9, 2026
  • docs(harnesses): record opencode's plugin-activation gate on session/new Measured on 2.2.5: the first `session/new` is answered with `ServiceUnavailableError` at +6s while Location plugins are still activating, and the same process answers a real session at ~80s. The probe used to read that refusal as final and report a booting agent as blocked. Records the fix's three deliberate properties (same process, opencode only, bounded with a stated reason) and the measurements behind them.

    @rajebdev rajebdev committed Nov 8, 2026