Skip to content

feat: add Freebuff TUI harness with login, auth check and model picker - #8200

Draft
milesatao-svg wants to merge 21 commits into
OrchestratorInc:mainfrom
milesatao-svg:codex/copy-pr-6553-freebuff
Draft

milesatao-svg wants to merge 21 commits into
OrchestratorInc:mainfrom
milesatao-svg:codex/copy-pr-6553-freebuff

Conversation

@milesatao-svg

@milesatao-svg milesatao-svg commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Current head status (1dc154ee): merged with main adf9f6a3. AO's private standing instructions are no longer sent as a visible first turn. Freebuff has no qualified private channel, so spawn and restore of any session carrying them (every worker/orchestrator) fail before writing Freebuff state, with STANDING_INSTRUCTIONS_UNSUPPORTED. Freebuff sessions therefore cannot start until a private channel is qualified with live evidence. Statements below about standing instructions in the first turn, and the historical runtime results, describe earlier heads. Receipt: see the latest PR comment.

Code changes: +2093 -94
Tests: +3099 -22
Others: +118 -4

Account migration from @nikhilachale to @milesatao-svg. This draft continues original PR #6553 at original head 86ecaa6454a8fdf47047206903e35eb4a27303e2. Original commits and commit authorship are preserved. Descriptions and evidence below include historical results with their original sources.

What

feat: add Freebuff TUI harness with login, auth check and model picker

Original PR description, preserved

Registers Freebuff as a Terminal UI harness. The compiled catalog was qualified against the public 0.2.30 source; local native validation used the installed 0.2.32 binary. This replaces the earlier guarded, unregistered qualification hold.

What works

Area Behaviour
Install Settings → Harnesses offers npm install, update and uninstall (freebuff). ao doctor probes freebuff --version.
Auth check Local credentials are configured, not provider-verified, when the default profile contains a token/keychain pointer or CODEBUFF_API_KEY is set. Missing credentials are unauthorized. Effective $FREEBUFF_CONFIG_DIR overrides the default ~/.config/manicode.
Login Log in to Freebuff runs freebuff login (browser flow; the URL is also printed) in the harness auth terminal.
Models Freebuff's 12-row /model picker, with MiMo v2.6 Flash Fast as the default. Each row carries its own reasoning ladder (DeepSeek low/high/max, GPT-6 Luna low…xhigh, and so on).
Applying model/effort AO serializes startup per effective config directory, seeds the canonical model and migration stamp, then confirms the process-local model and effort through the running native picker. The guard remains held through first-turn acceptance. Settings preflight is side-effect free and null settings are handled safely.
Messages First turns, follow-ups and lifecycle notifications use one native bracketed paste at the final guarded terminal write. Semantic prompt history is kept free of terminal frames; embedded escape sequences are rejected. Empty Enter-only nudges and other harnesses retain their protocol.
Launch freebuff --cwd <worktree>. After confirming the current empty native composer, AO uses one native bracketed paste for the complete first turn. It reports successful delivery only after Freebuff saves the complete recent user turn, including its native pasted-text wrapper.
Standing instructions Freebuff has no system-prompt input, so AO sends them with the task as one first turn, the same approach as fx.
Activity Derived conservatively from the complete current native composer/footer and anchored running status. Transcript quotations, drafts and historical empty composers do not establish activity; styled cursor evidence is required for empty-input confirmation.
Restore freebuff --cwd <worktree> --continue <chat-id>. Restore lookup and update-when-idle preflight use the effective session environment, including project FREEBUFF_CONFIG_DIR. The chat is matched by the projectRoot recorded in its run-state.json, because worktrees sharing a basename share a chats dir. With no match AO starts fresh rather than letting --continue pick an unrelated chat.
Registration Domain enum, registry, sqlite CHECK migration plus schema repair, API enums (openapi.yaml, schema.ts), avatar (official Freebuff mark), product-ui label, docs (docs/harnesses/freebuff.md, agents table, spawn skill).

Not supported

Chat (ACP) mode, reviewer sessions, per-session tool allow/deny lists (Freebuff has no such input), logout from Settings (Freebuff only has an in-TUI /logout), and unattended terminal refresh on a new release.

Verification

Rebased head cecef0feb onto main at a95139bb (22 October 2026): merge conflicts are resolved; all 24 checks passed at this head. PR checks. The reconciliation preserves Hermes registration, provider-switch native session IDs and rollback handling, effective Freebuff environments, and Cursor/cloud task controls. Both generated API files were regenerated together. Two stale CI assertions now check the combined harness set.

Local Go 1.27.1 build/vet, complete Freebuff/Hermes/terminal-guard/domain/launch-helper/installer race suites, all 31 cases in SQLite race shard 3, and five Freebuff/Cursor task-form regressions passed. The complete local Manager race run failed two unchanged main-branch tests (a three-second transition wait and a missing macOS tmux prerequisite); both isolated retests passed with the existing bundled tmux configured only for that test process. The complete run is not counted as a pass.

Broader local frontend and TypeScript runs reached their time limits without a complete result. Full frontend/type checks, platform jobs, Docker, pinned lint and sqlc drift are verified in CI rather than claimed as local passes. The requested AO automated review could not start because its legacy system-tmux runtime was unavailable; it produced no verdict.

Runtime results below predate this rebase and do not qualify the rebased head. A later isolated instrumented a775 + local diagnostic patch run produced the exact real assistant reply FREEBUFF_IS_WORKING with full actual-send payload-hash correspondence and zero manual Enter; the diagnostic patch was never published. The reviewed recording shows Starting, with the reply verified separately afterward; the safe on-screen reply crop was unavailable. Exact source and evidence chronology.

Historical follow-up a775cdb1d (application code equivalent to 452adfbb9): all 24 checks passed at that head, with none pending or failed. PR checks.

Historical fresh-profile unchanged-head UI Create retest: failed; working chat was unresolved at that point. After inspecting Nikhil’s latest commit and separate-profile successful screenshot/manual report, one real Scratchpad → Freebuff → default MiMo v2.6 Flash Fast/no effort Create used a fresh isolated AO database/Electron profile and the existing approved native config in place. POST /api/v1/sessions returned 409 SPAWN_DELIVER_PROMPT_FAILED after 39.184s; no assistant reply was saved. The reviewed 41.962s/117-frame actual renderer recording ends at Starting…, with the API error observed separately: the later error repaint was not captured. Inspected video, actual final still and full coverage.

The retained final native user/error snapshot has an inode birthtime 229ms after AO logged delivery failure, consistent with cancellation cleanup. Atomic replacement means this does not date the path’s first appearance or exclude earlier versions; missing optional sentAt does not establish the live failure cause. Complete expected combined-prompt equality is unproved because AO’s prompt file was removed; task/wrapper/length checks are not substituted. New full cwd/no --continue were verified; eight old history files and nine directory identities were preserved. No manual Enter, takeover, account/auth change or second case. The preceding native blocker remains unknown. At that point, no speculative acceptance change was claimed and the human working-Create request was still open. This is separate from Nikhil’s positive profile, the historical automatic reply below and the earlier recorded failure.

Earlier recorded native UI attempt (historical): failed. One actual Scratchpad standalone → Freebuff/default MiMo Fast Create click returned 409 SPAWN_DELIVER_PROMPT_FAILED after 43.956s, with no saved native user turn or assistant reply. The 44.970s actual Electron renderer recording shows the prepared harmless task form, click, wait and final visible error; all 154 decoded frames were visually reviewed before upload. Verified public video, decoded result still and exact coverage. This is a failed new-demo attempt, separate from the historical 02eae6289 automatic reply below. No manual terminal Enter or takeover was confirmed; original sessions and composer fingerprints were preserved. The video does not contain the native admission screen and does not establish a quota/takeover cause for this attempt. Earlier diagnostic 19 strongly matched the limited-account one-active-session gate through masked layout/source/binary evidence; its raw glyphs were not retained. At that point, the human takeover decision was pending and that attempt did not produce a working live demo.

At the previous head, all 24 remote checks passed on 02eae6289d5f15540bfc446d4fd682905797c7d9, including all four SQLite race shards, both core/chat race shards, session-store race, backend aggregate, frontend checks, Linux/macOS/Windows native coverage, container validation, API/sqlc drift and lint. The supplied follow-up review threads are resolved after their fixes were pushed.

  • Follow-up 452adfbb9 fixes a separate native long-paste retry blindspot: one current removable text attachment requires a focused idle composer, matching fresh paste notice, exact UTF-16 size and entire visible preview before a bounded Enter retry. General activity parsing and complete recent saved-user acceptance stay unchanged; unknown layouts and admission/takeover dialogs do not authorize Enter. The sanitized native fixture failed before the fix; complete Freebuff -race passed in 8.258s, actual-plugin-to-Manager saved-user integration in 2.503s, and full backend build, vet and pinned lint passed. The complete local Manager race suite also passed in 455.173s. The final metadata-only follow-up pins the native fixture to LF; checkout filtering with both core.autocrlf=false and true now preserves its exact bytes. This fixture-proven repair is not a runtime claim that the account takeover gate is cleared.
  • Migration follow-up: explicit model seeding now writes the exact native migration ID even when settings contain another nonempty stamp. New old/unknown/future-string cases failed before the fix and pass afterward; missing/current/blank/invalid stamps and byte-preservation without a model are covered. The complete Freebuff race suite passed in 7.487s, focused manager startup/restore/environment regressions in 2.615s, and full backend build, vet and pinned golangci-lint 2.13.2 passed with Go 1.27.1 (0 issues).
  • Previous effective-config follow-up: the complete session-manager race suite passed in 406.461s. Focused project-config seeding, chat lookup and spawn/restore environment regressions also passed under -race. Full backend build, vet and pinned golangci-lint 2.13.2 passed with Go 1.27.1 (0 issues). The new update-eligibility fixture reproduced both wrong outcomes before the fix and passed afterward: project-only native history allows updating; daemon-only history cannot authorize stopping a project-configured session. Earlier complete Freebuff, guard, daemon and ports race suites passed on 8be37e5a6.
  • The broader unsharded local backend run did not pass: SQLite exceeded its cumulative 20-minute package timeout. All other packages passed. The migration test running at that deadline passed separately under the race detector in 37.871s. Full local backend-suite success is not claimed; CI runs SQLite as four shards.
  • Local full frontend Vitest: 503 files, 8,574 passed, 6 skipped. Final-head CI: 503 files, 8,605 passed, 6 skipped; product UI 160/160 and Cloud client 29/29 also passed. Final main TypeScript check passed in 7m31s, E2E TypeScript in 11s, and guarded-navigation browser integration passed. The local main TypeScript attempt reached its bounded 20m18s active limit without diagnostics and is not counted as a local pass. Final frontend CI.
  • Reproduced and fixed a real TCP replay/input-loss race. Tests require actual PTY receipt before success, including failed/short writes, missing acknowledgments, cancellation and a 128 KiB message. Styled tmux capture excludes historical input.
  • Regression tests cover effective project config and restore, startup serialization, fresh default migration, native model/effort confirmation, Unicode display widths, conservative activity and exact recent whole-turn acceptance. Final follow-up tests verify multiline preservation, semantic metadata, fresh harness resolution under the input lease, safety-refusal priority and actual lifecycle-to-manager routing with exactly one paste frame.
  • Separate native 0.2.32 settings-only comparison: in two fresh isolated HOME/config directories, explicit DeepSeek plus a stale stamp migrated to MiMo; explicit DeepSeek plus the exact current stamp stayed DeepSeek before/after process stop. No credentials were copied and no input or inference was sent. This validates native stamp behavior independently of the AO adapter; the adapter write is covered by the regressions above.
  • Historical native automatic first turn passed on 02eae6289. In the real isolated Electron app with installed Freebuff 0.2.32, one standalone CLI spawn omitted model/effort and used the existing qualified MiMo Slow preference. It returned HTTP 201 in 13.190s and saved actual assistant text exactly FIRST_OK by 24.210s, with zero manual Enter. Independent inspection confirmed a recent same-workspace user record with the exact harmless task/unique marker in a 7,014-character [Pasted Text] turn. Complete expected combined-turn acceptance is enforced by the API; standing instructions were not independently reconstructed/hashed. Only the test session was cleaned up. The original authenticated chat, terminal-host PIDs and both composer fingerprints were preserved; the app remains open. No new native footer/model-label capture was made.
  • First-turn follow-up regressions passed under Go 1.27.1 -race: complete Freebuff 8.129s, shared Manager confirmation/failure/dialog/cleanup 2.193s, and input replay/ACK/cancellation plus actual macOS OS-PTY round trip 14.237s. Native-format saved-record fixtures, fake-PTY replay cases and real OS-PTY shell receipt are separate from the live provider reply above.
  • Earlier native installation/login/catalog checks and the automatic R3 MiMo Slow capacity denial remain historical evidence. The current source was also rebuilt/refreshed successfully before the new first-turn case: exact lab executable ready and native window visible. Effective-env update preflight remains verified by saved-chat regressions rather than a new native project-override restore.
  • The earlier 02eae6289 omitted-model automatic submission/response passed, but later current-02 visible UI/default and saved-Slow demo attempts failed as described above. No current-head successful reply, successful UI creation, resume, concurrent startup or native project-env restore is claimed. The new current-head UI failure is recorded above. The original manual-submit conversation/resume and earlier capacity-blocked attempts remain separate history; original authenticated sessions and drafts are preserved.
  • A separate harmless 3,131-byte native paste control showed a complete pending attachment and one Enter reaching the MiMo Fast quota gate. That is separate from the automatic task; no generated-prompt character count is claimed.
  • Local gaps: Docker is unavailable; container and Windows/Linux workflow coverage passed in final-head CI. Standard macOS predev could not compile the wake helper on SDK 13.3/Swift 5.8.1. Direct Forge ran successfully after rebuilding the unchanged pinned native dependency with an isolated Clang 18 toolchain; no source or dependency-version workaround was committed.

One binary avatar asset is included in the diff; its line count is zero.

🤖 Generated with Claude Code

Follow-up 28d92ed84: sets USERPROFILE alongside HOME in the cleared-config-override regression so Windows exercises the intended isolated home. The eight review findings already have fixes and resolved threads. The complete Freebuff adapter suite and focused Manager Freebuff/startup/after-start regressions pass locally on Windows/Go 1.27.1. Local race execution is unavailable because cgo is disabled; frontend dependencies are absent. New-head CI is pending; the preceding head's 24 checks passed. One binary asset has no line count.

Why

The contributor is moving this harness submission to @milesatao-svg. The linked source preserves the implementation and review history.

Before / After

Before After
Submission in PR #6553 from @nikhilachale at 86ecaa6454a8fdf47047206903e35eb4a27303e2 Submission from @milesatao-svg at the same commit 86ecaa6454a8fdf47047206903e35eb4a27303e2

Original product behavior and existing UI evidence are included in the preserved description and copied evidence comments.

How

Copy the source head into milesatao-svg/agent-orchestrator:codex/copy-pr-6553-freebuff and target upstream main.

Testing

Historical test results and screenshots are reproduced in the evidence comments with their original sources. Fresh validation is tracked by this PR's GitHub checks.

Checklist

  • Original head SHA and commit authorship preserved.
  • Original description and evidence provenance included.
  • Fresh CI validation on this submission.
  • Resolve current main merge conflicts.

Account migration verification

  • Exact source head copied: 86ecaa6454a8fdf47047206903e35eb4a27303e2.
  • New fork branch: milesatao-svg/agent-orchestrator:codex/copy-pr-6553-freebuff. Base: main.
  • Evidence comments retain their original author, timestamp, source URL and result wording.
  • Fresh validation status is tracked by this PR's GitHub checks.
  • Source mergeability at migration: conflicts with current main; reconciliation pending.

Ubuntu and others added 18 commits October 29, 2026 12:50
Turn the guarded, unregistered Freebuff preparation into a working harness.

- Launch `freebuff --cwd <worktree>` and type the first turn once the
  composer shows "Enter a coding task"; AO standing instructions travel in
  that turn because Freebuff has no system-prompt input.
- Auth: read the default profile of Freebuff's credentials.json (token or
  keychain pointer) or CODEBUFF_API_KEY; Settings > Harnesses offers
  `freebuff login`, and npm install/update/uninstall.
- Models: Freebuff's 12-row /model picker with per-model reasoning ladders.
  The choice is applied through settings.json (freebuffModel /
  freebuffReasoningEfforts), the same keys Freebuff's pickers persist.
- Restore: `--continue <chat-id>`, resolving the exact chat by the project
  root recorded in its run-state.json; no match falls back to a fresh launch.
- Terminal activity detection and supervisor exit detection (no hooks).
- Register the harness: domain enum, registry, sqlite CHECK migration and
  schema repair, API enums, doctor, avatar, product-ui label and docs.
- Replace the qualification-hold notes and evidence with docs/harnesses/freebuff.md.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…st turn

- Effort: route Freebuff through the catalog-validated effort gate (was
  rejected as UNSUPPORTED_EFFORT); inherited levels a model does not offer
  clear silently, and the task composer never sends a remembered level the
  selected model does not advertise.
- First turn: prime the empty composer with a lone Enter (Linux line-feed as
  Enter), then confirm from the screen that the turn left the composer,
  re-sending Enter at most three times only while the draft is visible, and
  fail the launch if it never submits. New optional ports
  AgentAfterStartSubmitConfirmer and AgentStartupReadyObserver.
- Model selection: write it in GetAgentHooks, which sees the session's real
  env on spawn, restore and switch, into the session's FREEBUFF_CONFIG_DIR;
  hold it with a per-directory lease until the TUI is ready so concurrent
  launches cannot swap models; stamp Freebuff's default-model migration;
  treat null/non-object settings as empty; drop a stale saved effort.
- Restore and session info search the config dir the session actually used,
  recorded under the AO data dir; command builders have no side effects.
- Auth: local credentials report configured, not authorized.
- Activity: match only composer/status-bar chrome, never transcript or draft.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…arting

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@milesatao-svg

Copy link
Copy Markdown
Contributor Author

Copied source evidence 1/1

Historical material from original PR #6553, copied for this account migration. Source PR head when copied: 86ecaa6454a8fdf47047206903e35eb4a27303e2. Each original comment below retains its own author, timestamp and source link. Its test results and screenshots describe the original run and commit. Fresh checks for this submission are shown on this PR.

Original comment by @nikhilachale on 2026-10-11T16:53:05Z

Original comment

Remaining batch handoff at immutable evidence 42803b62ab42644fcf2dd845d913a24cc8012275: aggregate candidate report, Freebuff qualification report.

Freebuff0.2.26: authorized native BYOK validation/edit/follow-up/activity/executing-child Escape cleanup controls passed, but independent private standing instructions failed on an ordinary fresh task. PR6553 remains draft and unregistered. AO UI screenshot unavailable because no AO session was admitted; no substitute screenshot or integration success claim. Blackbox/Warp/Rovo/Hermes ACP prerequisites remain explicit in the report. Completed Docker Agent/Mastra/CodeBuddy reports and their real AO images are preserved. Owned native processes were reaped; no production/merge/release/deployment changes.


Original comment by @nikhilachale on 2026-10-11T17:09:15Z

Original comment

Final batch evidence head f9da74d5938484d3205da63fe88aa6e0ed8c87d5: 5/5 GitHub Actions checks green, including 150/150 fixtures. Runner SHA256 8887aa2e7101b9e6d3950bfca08ffddf2e1e6c1c3f1d030c8dec2676443aecea is unchanged. No local routine suites were executed.

Final candidate report · Freebuff native hold

Current source heads: Docker Agent #6540, Mastra #6541 and CodeBuddy #6546 each23/23 green with actual AO Electron screenshot evidence. Stakpak #6550 and Pool #6551 each20/20 green but draft/unregistered for native external blockers. Freebuff #6553 source 9fbb97e559c245633554ee6545836af03f7348d7 has21/21 green but stays draft/unregistered: an ordinary fresh task did not consume the private standing instructions; passing prompted-context/edit/follow-up/executing-child-cancel controls remain separate. No AO screenshot is claimed for any hold.

Blackbox release metadata and official alternative endpoints still404; Warp private instructions/auth, Rovo site/token/credits and Hermes ACP instruction delivery/preservation remain unqualified. Existing Hermes TUI #6536 and OpenJarvis #6537 are preserved, each23/23 current-head checks green. Polygraph is excluded as a meta-orchestrator; connector scope is separate. Pool's alternate DeepSeek credential presence check found none in the authorized environment/helpers/50 native and audit configs; no secret was printed or imported.

Published image re-download verification matched all6 PNG hashes, and7 earlier integration comments were read back; current reports and this comment were read back too. Source worktrees, native profiles and prior evidence are preserved. Owned Freebuff TUI/broker descendants are absent, protected PID625660 remains, production untouched. No merge/release/deploy.


Original comment by @nikhilachale on 2026-10-11T21:59:02Z

Original comment

Freebuff recovery source fcdc3db2099c20854647b59173200a7524bb32fe; evidence 12601e986650596de963ca71325887dec4bf7645.

Recovery report · source diagnosis · unapplied upstream patch.

Independently confirmed the released binary root-only instruction filter and retained native-005 system prompt: root rules present, private nested rules excluded, ordinary artifact present and private receipt absent. All original controls/failures are retained. Alternatives inspected include CLI/settings/profile scope, custom-agent resolver, internal transport and on-demand skills. Stock0.2.26 exposes no supported private additive input preserving defaults.

The patch provides a concrete proposed bounded private-file loader and additive SDK transport, with unexecuted loader regressions and explicit native acceptance criteria. It was not applied to native code or submitted upstream; its env variable is not a stock feature. Remaining prerequisite is an official released route, followed by native and actual AO proof. PR6553 stays draft/unregistered. AO UI screenshot: unavailable because mandatory standing-instruction admission failed; AO lifecycle/Kill/UI remain NOT_RUN. No local routine tests/builds or new native turns ran. Current-head CI is pending.


Original comment by @nikhilachale on 2026-10-11T22:17:44Z

Original comment

Freebuff recovery final receipt.

Source PR6553 fcdc3db2099c20854647b59173200a7524bb32fe: 20/20 green exact-head checks. Evidence PR6509 aa0a700d5df4d807aa09473ff7b73609777914cd: 5/5 green, including 150/150 audit fixtures. Passing fixtures; first149/150 failure retained. The short-window fixture now waits for its actual child marker;80ms/120ms bounds and production runner remain unchanged. Installed runner SHA8887aa2e remains unchanged.

Immutable recovery report · precise source diagnosis · unapplied source patch/regression proposal. Published report download/hash and PR bodies were read back and verified.

Remaining external prerequisite: stock Freebuff0.2.26 filters nested private instructions out of its system prompt and exposes no supported additive per-session input preserving native defaults/user rules. The proposed env input does not exist in this release and was not executed or published upstream. Need an official released route, followed by independent native instruction/cancellation proof and actual AO lifecycle/UI qualification. No duplicate PR, production registration or gate weakening.

PR6553 remains draft/unregistered. AO runtime/Kill/UI: NOT_RUN. AO PNG: unavailable because the mandatory independent instruction gate failed; no synthetic substitute or integration success claim. All original negative evidence/controls are preserved. No local routine suites, builds, new native turns, AO sessions or cleanup kills were performed. Both owned source checkouts are clean; production/native profiles/protected audit process were untouched.


Original comment by @nikhilachale on 2026-10-12T03:41:05Z

Original comment

Failure report and fix implementation started (2026-10-12).

Stock Freebuff 0.2.26 discovered the nested private instruction file but did not consume it as standing context on an ordinary task. The released system-prompt filter includes only root agents.md/claude.md; the private nested key was absent from effective system context. Explicitly prompting a file read was only a control, not a passing instruction-delivery test. Preserved diagnosis and proposal.

The previous patch was unapplied. Actual native source implementation is now scheduled in the active durable VPS fix worker, after Stakpak. It will add an explicit additive context input preserving native/project/user instructions, review fresh-session versus cached-context behavior, and cover real CLI/client/model-request composition in CI. The helper-only proposal is not treated as a complete fix.

A concrete upstream fix PR and exact-head CI/native results will be linked here. Public-source prerelease artifacts will be labelled separately from the private npm release; no claim that stock 0.2.26 is repaired. AO screenshots remain unavailable until the fixed native path is qualified and the AO integration is exercised. No Mac changes/tests/builds, no merge or release.


Original comment by @nikhilachale on 2026-10-12T07:29:22Z

Original comment

Implemented and published native Freebuff fix PR1505. This is actual public source and regression tests, not the previous unapplied patch. Exact head5712d04bc916125d17da46548874a0388dcc526e; CI-built Linux0.0.0-ci binary SHA 5071e80f5eec7dd7a80d055e0ca0641b17470c7da44067d5ba8f032af9108822. No stock npm0.2.26 repair or registered AO harness claim.

The explicit creation-time FREEBUFF_APPEND_INSTRUCTIONS_FILE/SDK appendInstructions snapshot preserves native/project/user defaults. Existing constructor forwarding works. SDK validates type/Unicode/bounds; POSIX file errors are fail-closed and sanitized. Subsequent turns retain the snapshot; changed/retrofitted history fails before inference. Cached compaction prompts are not falsely refreshed. Ordinary native history/traces retain context; custom agents without the knowledge placeholder and Windows file input are outside qualified scope.

Exact-head fork CI:5/5 focused regressions PASS, real constructor→run→native prompt→provider composition, original-run.ts negative control PASS, Linux runtime build and binary smoke PASS. Full SDK declaration checkFAIL, independently reproduced on unmodified main; final CI step explicitly preserves the failure. Upstream Public CIACTION_REQUIRED needs maintainer approval. Both native1505 and AO6553 remain draft. No routine local suites/install/builds were run; CI owns validation.

Actual exact-head native proof ran in the real AO Electron standalone terminal on the VPS: ordinary leading-dash task once, workspace edit, independent private/project receipts, AGENTS preservation, second turn after deleting the private receipt and changing the source file, retained creation snapshot, armed Escape with executing PID/kernel witness and no delayed effect, same native identity and third turn, then independent AO terminal close204. Escape5/5 and close7/7 controlsPASS. This is not registered manager/session Kill or native restore/Chat/auth/catalog conformance.

Immutable detailed report and curated receipts. All8 image URLs downloaded and SHAverified after publication, inspected for credentials. AO backend/UI fb60bd4097ebd4109bbe7ee68e908f722e6ee03f, clean VCS; actual servingPID994173/kernel83308714/executable SHA b30f2dfa7b1f1c796c5772cb96d7d5c32b4bfb53c39f2c4ff7ca669f2cca8ded before/after capture. Runfile ~/.ao/audits/native-harness-fixes-20261012/reports/codebuff/ao-shell-001/running.json. Runtime handleptyhost-v1:shellterm-46b259221d611db5, native leader997948/start83420715, modelGLM-5.3-Flash/Z.ai. Native identity digest 98a00a64ba65da215f59b2d0d2daa7490a0a07bb64fd3955a47c6d85489b4723; standalone shells have no managed agent launch generation. Images are scrot captures on the actual owned Electron/Xvfb display.

UTC2026-10-12T07:03:36Z, initial ordinary task settled, PNG SHA 0863bf0efda301a34b1ff1ea94c5aa06108d1a7261a26a76e8301c4acf8a44b3:

Unreleased5712d04 independent private/project instructions in actual AO standalone terminal

UTC2026-10-12T07:13:39Z, same native identity and third turn after Escape, PNG SHA bfdb7b8bf171dbd44e9fc02431e0468573b5b5747487cb87b4c23002e0b2e38b:

Unreleased5712d04 same native conversation after cancellation and continuation in real AO

Escape childPID999673/kernel83474776, marker2026-10-12T07:09:55.107636+00:00, Escape2026-10-12T07:09:55.948115+00:00, absentwithin5s/after25s, no delayed artifact, unrelated controlPID994833/start83338796/heartbeat continued. Independent close204 at2026-10-12T07:15:45.615164+00:00 reaped native leader997948 and child1000643, no delayed artifact. Actual pre-Escape execution screen and pre-close execution screen have full capture records in the report.

Failures retained separately: stock nested knowledge exclusion; nonUUID audit fixture before provider; automated evidence driver143 with origin unestablished; grouped paste left an exact owned draft; secondary-before-primary color detection; and manual25s cancel window missed, original late artifact preserved (failed screen). Later armed control kept the same delay/timeout/absence assertions. No global strict auditPASS inferred. Original stock AO screenshot unavailable because admission never created a session.

Cleanup: native/runtime/control closes204, list200/empty, daemon shutdown202; exact owned native/control/PTY hosts and Electron main993952 stopped, mux helpers stopped, compute lock released. Production625660/ao-self-hosted and source/sharedgit/profiles/credentials/prior evidence untouched. Only a verified regenerable old CI artifact was reclaimed after its native PID was absent. Runner SHA8887aa2e... unchanged; current6509 evidence-head CI is tracked separately. No merge, release, deployment, signup or billing change.

Final evidence CI: exact6509 head384c80e5d483fb7edd345148c9ca0cbdbf53fc17 is 5/5SUCCESS, including 150/150fixtures (fixture run, frontend/mac-helper run). Runner SHA8887aa2e... remains unchanged. AO6553 headfcdc3db remains20/20green and draft/unregistered. Native1505 full SDK declarations stillFAIL and upstream Public CIACTION_REQUIRED; no native full-green/released-support claim. Stage result is implemented/native-qualified but blocked on those external prerequisites.


Original comment by @nikhilachale on 2026-10-12T16:51:24Z

Original comment

Harness-specific reports, JSON receipts, screenshots and provenance now live on this integration PR branch, rather than the skill PR #6509.

Evidence index with embedded screenshots and JSON links.

These screenshots show native/prerelease controls in AO standalone terminals. The harness remains gated/unregistered; this is NOT registered AO integration or manager Kill/restore proof.

This is evidence relocation only. Existing tested source/native hashes and failures are preserved; the documentation commit is not a fresh runtime pass. Applicable checks run in PR CI.

Recorded harness evidence — see scope above


Original comment by @illegalcall on 2026-10-21T00:23:58Z

Original comment

The eight findings in this changes-requested review are addressed by commits already pushed to this branch. Current head is 02eae6289d5f15540bfc446d4fd682905797c7d9.

Finding Pushed fix
Concurrent launches selecting another session's model 61eb46c7c: per-effective-config startup guard spans seeding, native process-local picker confirmation and first-turn acceptance.
Project config override lost at launch/restore 61eb46c7c propagates effective session environment; a4e05efc8 also fixes update-preflight chat lookup.
Explicit DeepSeek replaced by default migration 02eae6289: explicit picks always receive the exact native migration stamp, including noncurrent saved stamps.
Transcript/draft text becoming authoritative activity 61eb46c7c: require complete current composer/footer chrome and anchored adjacent status; no transcript-tail substring scan.
Local credentials reported as verified c688f76c6: tokens, keychain pointers and API keys report configured; readiness preserves unverified status.
JSON null settings panic c688f76c6 handles empty settings; 61eb46c7c provides the current explicit nil-map normalization and guarded-startup regression.
Shared effort gate and missing UI handling c688f76c6 + 61eb46c7c: Freebuff catalog validation, inherited/reset/default-model handling, catalog-driven effort control and valid request payloads.
Successful launch with an unsent first turn 61eb46c7c + 8be37e5a6: one native paste, PTY-write acknowledgment, exact recent durable user-turn confirmation, and bounded matching-draft retries; dialogs/unknown states cannot acknowledge submission.

All eight supplied inline threads are confirmed resolved. Relevant race regressions, actual Manager.Spawn boundary cases, real macOS OS-PTY round trip, and all three focused Freebuff TaskComposer UI cases passed. All 24 CI checks pass on the current head, including the complete frontend suite (8,605 passed, 6 skipped), both typechecks and guarded navigation.

Current-head live first-turn integration also passed in isolated Electron with Freebuff 0.2.32: one API-created standalone task returned HTTP 201 in 13.190s and saved actual assistant text exactly FIRST_OK by 24.210s, with no manual Enter. It used the existing qualified MiMo Slow preference with model/effort omitted; no new native footer/model label was captured. Independent saved-history inspection confirmed the same-workspace task/marker and pasted-text wrapper; complete expected combined-turn comparison is enforced by the API. Only the throwaway test session was cleaned up; original authentication, chat, terminal hosts and drafts were preserved. This supersedes the earlier quota blocker for that single automatic first-turn case.

The live case was API-created: literal final UI Create, exact resume, concurrent native startup and native project-env restore remain unclaimed on this head. The original changes-requested review verdict remains for the reviewer to reassess; no approval or dismissal has been applied.


Original comment by @illegalcall on 2026-10-21T00:25:52Z

Original comment

Follow-up to this review: the shared Freebuff effort blocker is addressed by pushed commits c688f76c6 and 61eb46c7c, retained by current head 02eae6289. Manager now permits and catalog-validates Freebuff effort, preserves compatible inherited levels, handles explicit resets/default models, and clears incompatible inheritance. TaskComposer exposes model-supported effort controls, replaces stale levels and sends an explicit empty reset where needed.

The resolved effort thread records actual Manager.Spawn regression coverage and the rendered Claude/Opus/High → Freebuff/GLM → effort selection → Create request test. The focused backend race checks passed (2.573s), and all three Freebuff UI cases passed (6.08s). All 24 CI checks remain green on the current pushed head, including the full frontend suite.

The separate initial-submission P1 now also has a fresh automatic native response on this head: Freebuff 0.2.32 returned HTTP 201 in 13.190s and saved an actual FIRST_OK reply by 24.210s without manual Enter. That case was API-created with model/effort omitted; it is not a literal final-head GLM/High UI Create pass.

All eight supplied inline threads are confirmed resolved; the consolidated response maps each finding to its pushed fix. No additional source change is needed for this supplied review. Its changes-requested verdict remains for reviewer reassessment.


Original comment by @illegalcall on 2026-10-21T00:27:40Z

Original comment

Follow-up to the completed native validation review: both runtime P1s and all six source findings are addressed by fixes already pushed through current head 02eae6289d5f15540bfc446d4fd682905797c7d9. The consolidated response maps all eight findings to their commits. All eight supplied inline threads are confirmed resolved.

The effort blocker is covered by shared Manager.Spawn and rendered TaskComposer regressions: Freebuff now passes the supported-harness gate, validates model-specific effort, preserves/clears inheritance appropriately, and exposes valid effort choices in the task form. The backend race cases and all three focused Freebuff UI cases passed.

The initial-submission P1 additionally has a new current-head native automatic reply pass, separate from the original manually submitted conversation: isolated Electron / Freebuff 0.2.32 returned HTTP 201 in 13.190s and saved actual assistant text exactly FIRST_OK by 24.210s, with zero manual Enter. The recent same-workspace pasted user turn/task marker was independently inspected; full expected combined-turn acceptance is enforced by the API. Only the new test session was cleaned up; original authentication, conversation, terminal hosts and composer contents were preserved.

All 24 CI checks pass on this pushed head, including the full frontend suite, both typechecks, all Go race shards, native platform checks and container/drift/lint coverage. The live case was API-created with model/effort omitted and the existing qualified MiMo Slow preference; no new native model-footer label was captured. Literal final-head UI Create, native resume/concurrency/project-env restore, standard local predev/wake-helper functionality and removal of the historical upstream cache/toolchain caveats are not claimed by that one first-turn pass. Earlier 0.2.30 source inspection and original 0.2.32 runtime evidence remain distinct historical records.

No further code change is needed for the supplied review. The changes-requested review verdict remains for reviewer reassessment.


Original comment by @nikhilachale on 2026-10-21T04:25:15Z

Original comment

Screenshot 2026-10-21 at 09 48 04

Original comment by @nikhilachale on 2026-10-21T04:37:40Z

Original comment

Manual end-to-end test of the current head (02eae6289) in the AO desktop app, run locally on macOS with Freebuff 0.2.32 and a fresh, isolated AO data folder. Everything below passed:

  • Install: Freebuff downloads and installs from AO. The Health tab reports it installed (v0.2.32, latest, npm).
  • Harness page (Settings → Harnesses → Freebuff): the Account, Models and Health tabs all work, including the model list, login, and check again.
  • New task: creating a task with Freebuff works.
  • Worker sessions: Freebuff runs correctly as a worker agent.
  • Orchestrator: Freebuff runs correctly as the orchestrator agent.

Original comment by @illegalcall on 2026-10-21T04:45:15Z

Original comment

Latest automatic chat check: the new standalone session answered FREEBUFF_IS_WORKING using MiMo v2.6 Flash Fast, with zero manual Enter or retry. This was a local instrumented diagnostic build: published a775cdb1d plus unpublished metadata-only patch 2e760b44…, daemon SHA 52fe7335…, and the retained a775-equivalent Electron frontend. The instrumentation changes no input, retry, success, deadline or cleanup decision, but adds local I/O overhead; this is not an unchanged-head runtime qualification or a newly pushed product fix. Nikhil’s latest commit and successful screenshot/report were checked before testing.

  • One normal visible Scratchpad → Freebuff → default MiMo v2.6 Flash Fast, no effort control → Create task. No second task, manual terminal Enter, takeover, account stop, auth change, purchase or model cycling.
  • Private hashes captured from the actual send variable prove that the fresh chat’s entire saved user turn equals the complete payload with the single native [Pasted Text] wrapper: 6,955 → 6,969 UTF-16 units. The exact new working directory and new chat identity were verified; all 21 preexisting history entries retained their bytes/identities.
  • A later bounded read-only check of that same chat found actual native role=assistant text exactly FREEBUFF_IS_WORKING, with native sentAt 12:37:57.160 UTC, 26.264 seconds after the actual send. The session remained nonterminated. This establishes a real reply, independently of an API status.
  • The attached recording preserves all 123 actual decoded frames, 44.509 seconds, 2640×1720, H.264 without audio. Every frame was visually privacy-reviewed. It shows the real Create click and Starting… only: the later reply is independently verified from the saved conversation, not shown in this video. The recorder did not retain the POST result, so no historical API status is inferred.
  • Source inspection of the installed Freebuff 0.2.32 bundle and cached upstream revision d193d4544d2355312e34de98ad85c2656259b575 establishes that its SDK emits error-shaped interruption checkpoints during live work. Its initial fallback user record can lack sentAt; the real user record later gains it. Therefore the earlier “session ended” checkpoint and missing timestamp did not establish an actual provider failure, and do not justify rejecting all such checkpoints or weakening the acceptance guard. A persisted checkpoint has no settled discriminator; SDK settlement is not claimed from the error shape or file time.

The working session is preserved. A later read-only view check verified that this same session was selected, but its strict privacy guard could not obtain a safe exact-reply crop without scrolling; no reply screenshot is claimed. No terminal input or navigation occurred. No diagnostic instrumentation is committed or pushed. The earlier unchanged-build failures below remain historical evidence with their original limitations; this successful instrumented run does not retroactively explain their cause.


Historical unchanged-build retest and earlier diagnostics:

Fresh-profile retest of unchanged a775cdb1d: Create still failed; a working new chat is not demonstrated. I first checked Nikhil’s latest commit and successful MiMo Fast screenshot/manual report; his separate-profile success remains separate evidence.

  • Setup: fresh isolated AO database, Electron profile and run file; existing approved Freebuff config used in place, without credential copying or login changes. Exact source/binary mapping was reverified: base e4e39b08 plus the complete committed a775cdb1d diff; application bytes match 452adfbb9. Native Freebuff 0.2.32, Electron 44.7.0.
  • One actual UI case: Scratchpad → Freebuff → default MiMo v2.6 Flash Fast, no effort control, then one normal Create click for Reply exactly FREEBUFF_IS_WORKING, then wait for my next message. Actual process argv used the new full working directory, with no --continue or chat-ID override.
  • Result: separately observed POST /api/v1/sessions returned 409 SPAWN_DELIVER_PROMPT_FAILED after 39.184 seconds; message: spawn standalone-1: deliver prompt: first-turn native acceptance could not be confirmed. No assistant reply was saved. The 41.962-second recording visibly ends at “Starting…”: it does not show the later error repaint. Controller flush/hold metadata is not proof of error pixels.
  • Postmortem distinction: the retained final native snapshot containing the pasted user payload and a session-ended error note has an inode birthtime 229 ms after AO logged delivery failure. Freebuff saves by atomic rename, so this does not establish when the path first appeared or whether an earlier version existed during confirmation. Its user records omit optional sentAt, consistent with the native SDK’s cancellation fallback. This is not proof that a saved turn existed during live confirmation, or that missing timestamps caused this failure. The complete expected combined-prompt equality cannot be independently established because AO’s generated prompt file was removed during cleanup; task/wrapper/length checks do not replace that proof. The preceding native blocker remains unknown.
  • Limits/preservation: no manual terminal Enter, takeover, provider-session stop, purchase, auth change, model cycling or second case. All eight preexisting history files and nine directory identities remained unchanged. A new timestamp chat directory inside the shared basename bucket is expected. No speculative acceptance change was applied. All 24 checks remain green on a775cdb1d.

The attached MP4 preserves all 117 actual frames, their capture timing and native 2640×1720 raster, H.264 without audio. Every decoded frame was visually reviewed for privacy. The still is the actual final decoded Starting frame, not a manufactured error/result. Recording duration and API latency are distinct. The earlier recording and diagnostics below remain historical; none establishes the cause of this fresh failure.


Earlier recorded run and diagnostics (before the fresh-profile retest; historical process-preservation statements predate a host reboot):

Recorded native UI reproduction on a775cdb1d: FAILED; no new working demo. The attached 44.970-second video is actual Electron renderer content at native Retina resolution, beginning at the prepared harmless task form and showing one Create task click, the wait, and the final visible delivery error. It is not an OS desktop recording or an install/login recording.

  • Tested source: isolated base e4e39b08 plus the complete committed a775cdb1d diff, verified byte-for-byte/reverse-apply; running application bytes are equivalent to 452adfbb9 (a775 adds only the fixture LF attribute). Installed native Freebuff 0.2.32, Electron 44.7.0.
  • Actual UI path: Scratchpad standalone → Freebuff → default MiMo v2.6 Flash Fast, with no effort control. Exact task: Reply exactly FREEBUFF_IS_WORKING, then wait for my next message.
  • Observed result: POST /api/v1/sessions returned 409 SPAWN_DELIVER_PROMPT_FAILED after 43.956 seconds (daemon request log). The visible error is spawn standalone-20: deliver prompt: first-turn native acceptance could not be confirmed. No native saved user turn or assistant reply was found; the failed throwaway session exited automatically. The recording length is separate from the API latency.
  • Limits: no manual terminal Enter, takeover confirmation, provider-session stop, account/auth change or purchase. Original sessions, host PIDs and both composer fingerprints were preserved. This video does not expose the native admission screen and does not establish a quota/takeover cause for this attempt; the earlier diagnostic below remains separate evidence.

All 24 remote checks pass on a775cdb1d. The MP4 was independently decoded and every frame visually reviewed before upload; all 154 real captured frames and their timing were preserved. The final still is decoded from that same video for a readable result view. The human takeover decision remains unanswered; a fresh successful reply is not claimed.

Earlier diagnostics and fix validation follow; they do not replace the recorded failure above.

Update on the fresh-demo failures and follow-up commit 452adfbb9:

  • The actual UI default MiMo Fast attempt failed (409, no saved user turn/reply); a separate saved-Slow attempt also failed. These remain distinct from the earlier historical FIRST_OK pass.
  • A targeted current-02 diagnostic reached a panel whose complete redacted character/spacing pattern matches the installed 0.2.32 limited-account one-active-session takeover gate, including Enter: take over and Esc: back to draft. This was independently checked against native binary strings and the pinned native source. Original raw glyphs were not retained; this is strong source/pattern corroboration, not a recovered screenshot transcript. Which other provider instance holds the slot is unknown. The grey card above is consistent with native ad-card rendering, rather than the admission controls below.
  • No takeover was confirmed, since that stops another active Freebuff instance. The human has been asked to choose; no successful new live demo is claimed.
  • Separately, the pushed fix recognizes one fresh native long-paste attachment by its focused empty composer, single removable card, expected UTF-16 count, entire visible preview and current paste notice. It permits a bounded Enter retry only; success still requires the complete recent native user record. It does not approve takeover/admission dialogs. The native-format fixture failed before the fix. Complete Freebuff -race passed in 8.258s; actual-plugin → Manager → retry → saved-user regression passed in 2.503s; full build, vet and pinned lint passed. The complete Manager race suite also passed in 455.173s. Follow-up a775cdb1d pins the native capture to LF for Windows checkouts; paired checkout-filter checks preserve the exact fixture bytes, and application code is identical to 452adfbb9. All 24 remote checks passed on exact current head a775cdb1d, with none pending or failed: checks.

The authenticated dev app/original sessions and drafts are preserved. Fresh demo completion remains pending the takeover decision, rather than being reported as passed.

freebuff-a775-ui-create.mp4

Final decoded frame of the recorded Freebuff Create-task failure

freebuff-baseline-a775-ui-create.mp4

Final actual frame of the fresh Freebuff baseline: Starting, not an error repaint

freebuff-instrumented-first-create.mp4

@coderabbitai

coderabbitai Bot commented Nov 12, 2026

Copy link
Copy Markdown

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

@github-actions

Copy link
Copy Markdown
Contributor

🏆 Review leaderboard

Nov 5–12, 2026 · UTC

Rank Reviewer PRs reviewed Review rounds PR comments
🥇 @ronishrohan @ronishrohan 91 165 14
🥈 @neversettle17-101 @neversettle17-101 11 15 4
🥉 @illegalcall @illegalcall 11 12 4
4 @Meraj-08 @Meraj-08 10 15 0
5 @nikhilachale @nikhilachale 7 19 3
6 @Prasad-D-Ware @Prasad-D-Ware 7 12 2
7 @Annieeeee11 @Annieeeee11 6 9 0
8 @mohakchakraborty2004 @mohakchakraborty2004 6 7 2
9 @Pritom14 @Pritom14 5 8 1
10 @Rishet11 @Rishet11 5 6 3

Ranked by distinct external PRs reviewed, then review rounds, then PR comments. Self-activity and bot activity are excluded.

Show 6 more reviewers
Rank Reviewer PRs reviewed Review rounds PR comments
11 @somewherelostt @somewherelostt 3 11 1
12 @ApexYash11 @ApexYash11 2 6 2
13 @Vaibhaav-Tiwari @Vaibhaav-Tiwari 2 3 0
14 @codebanditssss @codebanditssss 2 2 8
15 @harshitsinghbhandari @harshitsinghbhandari 2 2 0
16 @MohamedKamran @MohamedKamran 1 2 0

@milesatao-svg

Copy link
Copy Markdown
Contributor Author

Account migration verification

  • New PR author verified as milesatao-svg; submission is a draft.
  • Original PR #6553 and this PR both point to 86ecaa6454a8fdf47047206903e35eb4a27303e2.
  • Published change-count header verified against this PR's actual GitHub file diff.
  • All 13 original human comment bodies and source links reproduced across 1 historical-evidence comments.
  • All 9 original screenshot/recording URLs reproduced and checked as HTTP 200 media.
  • Current GitHub mergeability: conflicts with main; resolution pending.
  • Fresh CI validation and current-head harness certification are pending. Copied test outcomes retain their original run/commit provenance.

@OrcInc OrcInc added comp/daemon Go daemon, process lifecycle, and backend control plane. enhancement New feature or request labels Nov 12, 2026
@OrcInc OrcInc added this to the Agents & orchestration milestone Nov 12, 2026
nikhilachale and others added 2 commits November 14, 2026 23:20
Resolve conflicts against main adf9f6a while keeping main's daemon,
API and UI behavior:

- Session manager: adopt main's freshLaunchArgv/restoreArgv signatures
  (trailing env), launchPath and nativeRestoreMetadata; keep the
  effective env on native restore and interface-transition preflight
  (harness-scoped, side-effect free); drop LaunchConfig/RestoreConfig
  Env keys that both sides added.
- Effort: keep main's Grok Chat clause and spawn model-check step;
  allow Freebuff TUI effort alongside it.
- TaskComposer: keep main's cloud effort logic and TUI-fallback hiding;
  scope Freebuff's stale-rung, inherited-effort clearing and TUI
  effort behavior to Freebuff only.
- Settings: route Freebuff's configured login through
  authWorkflowSucceeded; keep main's Claude test beside the
  parameterized MiMo/Freebuff test.
- Renumber the Freebuff harness migration above main's latest
  (20261114172624733) and move its ledger entry.
- Register Freebuff's CODEBUFF_API_KEY provider method and agent tone,
  required by main's all-harness checks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nikhilachale

Copy link
Copy Markdown
Collaborator

Merge with main + ao-agent-harness receipt (head 82b00eb4)

Head: 82b00eb4d0aaaf368638c4d8254b8cd38f82bf93: merge of main@adf9f6a3 into 86ecaa64, plus one docs commit. Fast-forward push; no rebase or force-push.

Conflict resolution (main's behavior kept)

  • Session manager: main's freshLaunchArgv/restoreArgv signatures (env last), launchPath, and nativeRestoreMetadata. Keeps the PR's effective env on native restore and in interface-transition preflight (now runtimeEnvForHarness, no capability issuance). The startup guard now uses launchPath.
  • Merge-only compile fixes: both sides added Env. Removed the duplicate keys in spawn/restore/fresh LaunchConfig and the duplicate RestoreConfig.Env field.
  • Effort: main's Grok-Chat clause and spawn model-check step are kept. Freebuff is added to the TUI effort allowlist.
  • TaskComposer: main's cloud effort logic and TUI-fallback effort hiding are kept. The PR applied three rules to every harness: drop stale rung, clear incompatible inherited effort, show effort in TUI fallback. They are now scoped to Freebuff. Otherwise, custom/uncatalogued models would lose saved effort, and Grok's TUI fallback would show an effort it rejects.
  • Settings: Freebuff joins main's authWorkflowSucceeded. Main's Claude test sits beside the parameterized MiMo/Freebuff test.
  • Migration: renumbered 20261020181743263 to 20261114172624733 (above main's latest), with its ledger entry moved. The SQL is idempotent, and WithAllowMissing plus reconcileHarnessConstraint still cover DBs at older versions. No later main migration rebuilds the sessions harness CHECK.
  • Main's all-harness checks: added env_CODEBUFF_API_KEY to providerEnvironment (the adapter already accepts it, same pattern as Amp) and a neutral agent-tone entry, which the Record<AgentId,…> requires.

Skill receipt (static, from 04c8c95b; scripts not run on this host)

Step Result Evidence / gap
0 Intake PASS This is the open PR for Freebuff; continued, not duplicated
1 Compatibility PASS (doc) docs/harnesses/freebuff.md covers install, login, TUI, restore, models, unsupported (Chat, reviewer, tool lists, logout)
2 Touchpoints PASS on all required items except one N/A activity-dispatch is N/A: Freebuff has no hooks, documented, uses TerminalActivityDetector. Optional gaps: avatar licence file
3 Contract SKIP No local-contracts.json entry written; contract and run artifacts are not committed
4 Live audit BLOCKED No execution on this Windows host by user rule; needs a Freebuff account (below)
5 UI audit BLOCKED Same as step 4
6 Hand-off PARTIAL Items below

Guide "Avoid" findings

  • FAIL, decision pending: standing instructions go in the visible first turn (## AO Standing Instructions), following the fx precedent, because Freebuff has no system-prompt input. The guide forbids this. Options: accept it as a documented deviation, add a session-scoped rule file (needs live proof Freebuff loads it), or fail clearly.
  • Auth: reports configured only. There is no upstream verification probe, so "Connected" (step 6 of the UI audit) is unproven.
  • First turn: bracketed paste; success only when Freebuff's run-state.json saves the complete user turn after this startup (projectRoot-matched, [Pasted Text] wrapper accepted). Enter retries are bounded (3) behind a settled matching draft. The startup guard is held through acceptance, and failure gives SPAWN_DELIVER_PROMPT_FAILED with rollback. Gap: native admission/takeover dialogs surface as that generic 409, not a typed needs-user error.
  • Shared code: tmux GetStyledOutput is viewport-only, and conpty input waits for a PTY ack. Main has added more styled-output callers since (shell-agent detection, shell-terminal auth, admission, activity observer). These still get bounded recent rows, but there are no Codex/Claude-specific regression tests for this change.

Validation

  • No local builds or tests (Windows host rule). Statically checked: no conflict markers, no duplicate struct keys in Go files changed on both sides, no new duplicate functions versus main, LF endings, enum parity (hermes/freebuff 3/3 in openapi.yaml and schema.ts).
  • CI has not run on 9329181b or 82b00eb4. The fork workflows are action_required and need maintainer approval.

Live testing needed (user)

  • A Freebuff account signed in through freebuff login (or CODEBUFF_API_KEY). MiniMax/Z.ai keys cannot exercise Freebuff. A limited account's one-active-session gate previously blocked the first turn.
  • Spawn with default MiMo, plus one effort-capable model. Check the native saved user turn, not just HTTP 201. Then follow-up, Kill/restore (--continue <chat-id>), and a project FREEBUFF_CONFIG_DIR restore.

Not merge-ready: needs CI approval and a green run, fresh review, maintainer approval, the system-prompt decision, and live/UI proof.

…g them

Freebuff has no system-prompt input and AO has not qualified another
private channel, so AO's standing instructions were sent as a visible
first user turn. Stop doing that: the first turn now carries only the
task, and a session that carries standing instructions fails with
ports.ErrStandingInstructionsUnsupported (STANDING_INSTRUCTIONS_UNSUPPORTED,
4xx, treated as a config failure by the auto-spawn breaker).

The refusal runs at every adapter entry point (delivery strategy, first
turn, startup guard, launch, restore) before any Freebuff config or
settings are written. Restore now passes the session's standing
instructions to the startup guard so it can refuse there too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nikhilachale

Copy link
Copy Markdown
Collaborator

Receipt update: head 1dc154ee (decision (c))

Change: Freebuff no longer pastes AO's standing instructions into the conversation. BuildAfterStartPrompt returns only the task. A session carrying standing instructions (inline or file) is refused with ports.ErrStandingInstructionsUnsupported, which maps to STANDING_INSTRUCTIONS_UNSUPPORTED (KindInvalid) and is listed as a config failure in the auto-spawn breaker. The refusal runs first in GetPromptDeliveryStrategy, BuildAfterStartPrompt, BeginStartup, GetLaunchCommand and GetRestoreCommand, so nothing is written to the Freebuff config directory or settings. Restore now passes the session's standing instructions into the startup guard so it refuses there as well. Docs and the spawn skill now say Freebuff sessions cannot start yet.

Effect: every AO worker and orchestrator session carries standing instructions (buildSystemPrompt always adds the role prompt), so Freebuff spawn and restore now always fail clearly. This is intended until a private channel is qualified.

Tests added (not run locally; Windows host): TestStandingInstructionsAreRefusedNotPasted covers all five entry points, inline and file-based instructions, and confirms the config directory is never created. TestBuildAfterStartPromptSendsOnlyTheTask. Two toAPIError cases cover the spawn-stage wrap and the restore RelaunchError wrap.

Skill item Status
System prompt: never paste private instructions PASS (refused, not pasted)
System prompt: qualified private channel BLOCKED: none qualified. A session-scoped rule-file candidate needs isolated, unregistered work plus live proof that Freebuff loads it; not started in this PR
First-prompt delivery / native saved turn Code path unchanged and statically verified. Unreachable live while the refusal holds
Step 4 live audit / step 5 UI audit BLOCKED: Windows host rule; Freebuff account needed; launches also refuse until a channel exists
CI on 9329181b, 82b00eb4, 1dc154ee BLOCKED: fork workflows are action_required and need maintainer approval

Still open: maintainer approval to run CI and a green run; fresh review; maintainer approval; a qualified private channel with live proof before Freebuff can launch at all; then a live test on a Freebuff account (MiniMax/Z.ai keys cannot test it). Earlier receipt items still apply: auth is configured-only, the native dialog errors are generic, and the shared tmux/conpty change has no Codex/Claude-specific tests. Not merge-ready.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

comp/daemon Go daemon, process lifecycle, and backend control plane. enhancement New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants