Bare "403" and "forbidden" substring patterns (added in 034744cb to
retry the "Selected provider is forbidden" case from PR #3706) matched
any error message containing those tokens — tool-level 403s, file-perm
"forbidden" messages, unrelated upstream errors that happened to spell
"forbidden". The legacy model-fallback path then armed setPendingModelFallback
on those unrelated errors, jumping Sisyphus to claude-opus-4-7 (first
entry of its fallback chain) regardless of the user's configured model.
Replace the bare patterns with the specific phrases PR #3706 actually
targeted, and add regression tests asserting unrelated 403/forbidden
messages stay non-retryable.
Reported-by: ilove_borshch on Discord (#omo-help)
- Add Manual QA Gate as a non-negotiable surface-tool mapping
(interactive_bash for TUI, playwright for browser, curl for HTTP,
driver script for library) to Hephaestus, Sisyphus-Junior, and
Sisyphus' direct-execution mode.
- Restore the dig-deeper trio (tool persistence / dig deeper /
dependency checks) as orthogonal paragraphs so each carries its
own cognitive trigger instead of a fused single statement.
- Harden investigate-before-acting from a soft phrase to a dedicated
block: never speculate about unread code, re-read on every
task hand-off, the worktree may have changed.
- Add 'Parallelize aggressively' as its own block. Reads, searches,
diagnostics, and background sub-agents all batch into a single
response by default.
- Add 'No defensive code, no speculative legacy' to discourage
speculative backward-compatibility branches and unrequested
defensive validation.
- Absorb review-intent and frontend anti-slop coverage so the
prompt stays self-sufficient when the omo agent prompt replaces
the provider prompt.
- Replace literal apply_patch instructions with GPT_APPLY_PATCH_GUIDANCE
(use edit/write tools) so the prompt no longer contradicts the
apply_patch deny that the agent permission applies on GPT models.
- Sisyphus-Junior gains a Review tasks block and a default-behavior
fallback for when the runtime category context is missing or sparse.
- Sisyphus gains an explicit Hard invariants block listing type-
suppression bans, destructive-git bans, and Oracle-completion gating.
- Restore dynamic injections that round out the orchestrator/worker
context: category+skills delegation guide, delegation table,
Oracle dynamic guidance, key triggers, non-Claude planner
reminder. Hephaestus regains optional category delegation while
keeping direct execution as the default.
- Drop em dashes; search guidance points at rg directly throughout.
When shutdown() itself emitted uncaughtException (e.g. EPIPE while closing
a broken pipe), the error listener re-entered itself, re-logged, re-ran
cleanup, and threw EPIPE again. The 6 s forced-exit timer could not fire
because every re-entry stalled the event loop with fresh synchronous work.
Users hit this after v3.17.5 and observed 100+ GB of log lines written to
disk within minutes, with one confirmed report of a 157 GB log file filling
the filesystem.
Detaching the listener with process.off() before running log() + handler()
breaks the loop at the first re-emit: the second event has no listener to
invoke, and the first invocation's scheduleForcedExit() proceeds normally.
The startup toast and `omo --version` were reading from the legacy flat install
at <CACHE_DIR>/node_modules/<pkg>/package.json, but OpenCode actually loads
plugins from a per-plugin sandbox at <CACHE_DIR>/<plugin-entry>/node_modules/
<pkg>/package.json. The two install layers can drift independently when bun
re-resolves "latest" against the flat install while the sandbox's package.json
stays pinned to a literal version baked in at first install.
In practice this means the toast can announce a version the runtime is not
running. Concrete reproduction: with `"oh-my-openagent@latest"` in the plugin
list, the sandbox stayed on 3.17.5 while the parallel flat install advanced to
3.17.6, so the startup toast confidently reported v3.17.6 even though the
loaded plugin code was 3.17.5.
Walking up from `import.meta.url` always reflects the actually-loaded module,
so reorder `getCachedVersion()` to try that first and fall back to the flat-
install candidates and execPath walk-up as before. The fallback chain is
preserved for bundled environments where the module-relative lookup may fail.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Hephaestus 5.5 was rewritten as an outcome-first delegation contract in c3fabaaf. The deep category (spawned as sisyphus-junior under gpt-5.5) now receives a matching prose-driven category context lifted from drafts/gpt-5-5/deep.md instead of the legacy gpt-5.4-era threat-frame version.
Selection happens via a new model-aware resolvePromptAppend hook on BuiltinCategoryDefinition. When the resolved category model is gpt-5.5 the new DEEP_CATEGORY_PROMPT_APPEND_GPT_5_5 is used; older models keep the legacy DEEP_CATEGORY_PROMPT_APPEND. User prompt_append remains preserved on top of either base.
Previously each transform hook was awaited sequentially without per-hook
error handling. If contextInjectorMessagesTransform or thinkingBlockValidator
threw, toolPairValidator was silently skipped, leaving orphaned tool_use
blocks in the post-compaction API payload and producing
"messages.N: tool_use ids were found without tool_result blocks immediately
after" 400s from Anthropic.
Wraps each hook in runHookSafely so an upstream throw is logged but the
chain continues. Adds regression tests covering the isolation contract and
the consecutive-assistants compaction tail case (ses_22bd806).
The idle notification scheduler was initialized with platform 'unsupported'
before platform detection had run. This stale value was passed to send/playSound
callbacks, causing sendSessionNotification to silently no-op (no switch case for
'unsupported'). Session-idle notifications never fired as a result.
Fix: remove platform from scheduler options entirely. Callbacks now resolve
platform via ensureNotificationPlatform() which is sync, cached, and already
called by each event handler before scheduling.
Add cmux as the highest-priority notification provider on macOS.
cmux delivers notifications via native UNUserNotificationCenter,
properly attributed to the cmux app instead of Script Editor.
Notification priority: cmux > terminal-notifier > osascript
Tests cover the full fallback chain:
- cmux available: uses cmux, skips others
- cmux fails: falls back to terminal-notifier
- cmux + terminal-notifier fail: falls back to osascript
- cmux not available: skips to terminal-notifier
Fixes#3628
Root cause: the context-window-monitor hook computes actualUsagePercentage = (input + cache.read) / actualLimit and renders both 'X% used' and '(1 - X) * 100% remaining' inside a [SYSTEM DIRECTIVE: OH-MY-OPENCODE - CONTEXT WINDOW MONITOR] block that is appended to bash tool output. When resolveActualContextLimit() underestimates the model's real context window (for example a 1M-context Anthropic model that falls back to the 200K default per #3450), totalInputTokens > actualLimit and the rendered numbers go nonsensical (issue #3655 reproduces 144.7% used / -44.7% remaining at 289,370 / 200,000 tokens). Safety-tuned models recognize the >100% / negative-remaining pattern as a tell-tale prompt injection and refuse to follow the directive.
Fix: clamp actualUsagePercentage to [0, 1] before formatting. The 70% threshold check still uses the raw value so the block continues to fire above threshold, and resolveActualContextLimit() is left untouched (the deeper resolver concern is tracked separately as #3450). When totalInputTokens exceeds actualLimit the displayed numbers now read '100.0% used / 0.0% remaining' instead of the impossible >100% / negative pair, and safety-tuned models stop flagging the block as an injection attempt.
Verification: added a regression test (input 289,370, limit 200,000) that asserts usedPct in [0,100] and remainingPct in [0,100]. Test fails before the fix (Received: 144.7) and passes after. Full context-window-monitor.test.ts and context-window-monitor.model-context-limits.test.ts: 15 pass / 0 fail. Typecheck clean.
The plugin_loaded event was emitted on every plugin reload, generating
high event volume on PostHog (proportional to opencode restarts per
user per day). With MAU > 60K and active power users restarting
frequently, this drove unnecessary event spend.
Add a separate daily dedup state field (lastPluginLoadedDayUTC) so the
plugin_loaded capture only fires once per UTC day per machine. The
existing daily activity dedup (lastActiveDayUTC, used by
omo_daily_active) is preserved as an independent gate so the two
dimensions cannot overwrite each other in the activity state file.
- New src/agents/sisyphus/kimi-k2-6.ts based on gpt-5-4.ts 8-block architecture
- New src/agents/sisyphus-junior/kimi-k2-6.ts worker variant
- Preserves 4 pillars: intent gate + verbalization, parallel tools, verification
- Adds <re_entry_rule>: suppress re-verbalization for already-resolved turns
- Adds <exploration_budget>: hard stop conditions alongside aggressive parallelism
- Tiered <verification_loop> (V1/V2/V3): V3 keeps full rigor with harsh enforcement
- Adds <token_economy>: verbalization explicitly excluded from trim mandate
- isKimiK2Model in types.ts: matches kimi, k2p5/k2p6 variants (case-insensitive)
- Routing in sisyphus.ts + sisyphus-junior/agent.ts
- Tests: 3 new kimi routing cases in sisyphus-junior/index.test.ts (all pass)
Motivation: K2.x was post-trained with Toggle RL (~25-30% token reduction) and a
GRM scoring appropriate detail + intent inference. Reusing Claude-style prompts
double-taxes the model — external strictness on top of RL-learned strictness causes
over-deliberation on already-resolved requests. The re-entry rule and exploration
budget fix this without weakening verification rigor.
Refs: kimi.com/blog/kimi-k2-6, arxiv 2602.02276 §4.4.2 (Toggle, GRM)
The Codex 5.2 restyle in ad9df3f68 watered down the four deep-work
exhortations from gpt-5.4 (tool_call_philosophy, tool_persistence,
dependency_checks, dig_deeper) into a single bullet, leaving the
'deep worker' identity without behavioral teeth.
Restore them as Codex-style sub-sections under Exploration:
- Tool-call discipline: more calls = more accuracy, retry on partial,
read more files than needed.
- Dig deeper: don't stop at first plausible answer, check second-order
issues, prefer root over symptom (with concrete example).
- Dependency checks: resolve prerequisites before acting.
- Anti-duplication: extracted from inline paragraph to its own block.
LSP clean. 267 -> 315 lines.
Previous prose-dense rewrite went too far in stripping bullet structure.
Codex 5.1/5.2 prompts (the closest reference for an OpenAI deep-worker
prompt) actually use bullets liberally - just well-grouped (4-6 per list)
with prose introductions on each section. Restructure 5.5 to mirror that
style and tone while preserving Hephaestus's identity and all behavioral
rules from the prior round.
Sections lifted directly from Codex 5.1/5.2 organization:
- # How you work / ## Personality at the top for tonal priming
- # AGENTS.md spec as a standalone section with its own bullets
- ## Autonomy and Persistence with prose intro + Three-attempt sub-protocol
- ## Responsiveness with Frequency, Tone, Content, Examples sub-blocks
(examples rewritten to Hephaestus voice: 'Walking the agents/ tree',
'Found the dispatch in createSisyphusAgent', etc.)
- ## Plan tool with 'use a plan when' bullet list
- ## Validating your work with approval-mode granularity
(non-interactive / interactive / test-related)
- ## Presenting your work with categorical Final answer rules
(Section Headers / Bullets / Monospace / File references / Tone /
Verbosity / Don't)
- # Tool Guidelines as separate top-level section
Hephaestus-specific content preserved verbatim:
- Forge god identity, deep-worker / executor framing
- task() restricted to research subagents only
- Three-attempt failure protocol
- End-to-end usage gate (interactive_bash / playwright / curl / driver)
- Anti-duplication rule on parallel exploration
Amp-derived rules kept compact in their own ## Pragmatism and Scope:
- Smallest correct change, duplication > premature abstraction
- Default-no-tests with explicit exceptions
- WIP-not-legacy rule
- Multi-agent dirty worktree safety
Metrics: 110 -> 267 lines (still -14% from original 312), 4 -> 100 bullets
(grouped Codex-style, not scattered), 24 headers. 38/38 verification
checks pass; LSP clean.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Prior 5.5 prompt enumerated rules across 60+ bullets and 312 lines, which
fights GPT-5.5's strength: it follows prose instructions reliably and does
not need rule-by-rule cataloging. Rewrite as flowing paragraphs while
preserving the deep-worker identity and every load-bearing behavior.
Identity preserved:
- Forge god mythology ("Your boulder is code", "forge it until done")
- Direct executor, not orchestrator (research subagents only)
- Senior-colleague tone, end-to-end persistence
Behaviors preserved (compressed to prose):
- Three-attempt failure protocol → 1 paragraph
- Anti-duplication on parallel exploration
- End-to-end usage gate (interactive_bash / playwright / curl / driver)
- Implementation gate: when delegated, execute directly, no draft loop
Net additions distilled from Amp + Codex 5.2 evolution:
- Pragmatism block: smallest correct change, duplication > premature
abstraction, do not over-engineer, do not validate impossible scenarios
- Default-no-tests: add tests only when user asks, fixes a subtle bug,
or protects an important boundary; never to codebases without tests
- WIP-not-legacy: earlier unreleased shapes in the same turn are drafts,
not legacy contracts requiring backward compatibility
- Multi-agent worktree: continue task without reverting unknown changes
- Code-review mode trigger: "review" → findings-first, severity-sorted
- Personality-first opener (Codex 5.2 pattern) for tonal priming
Metrics: 312 → 110 lines (-65%), 60+ bullets → 4 bullets, 21,803 → 15,654
chars (-28%). 26/26 verification checks pass.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Hephaestus is the autonomous deep-worker - everything it receives is a
delegation. Existing 'Manual behavior' bullet vaguely said 'actually run
it' but left the validation surface unspecified, which lets a checked-in
diff plus passing tests masquerade as completion on user-visible work.
Add a dedicated 'End-to-end usage is the gate' subsection in Codex prose
style (no threats/CAPS, contract frames). Surface determines tool:
- TUI / CLI → interactive_bash (tmux), drive it like a real user
- Web / browser / UI → playwright skill, drive a real browser session
- HTTP API / service → curl or integration script against running service
- Library / SDK → minimal driver script
Reinforce in Forbidden stops trailer: when receiving a delegation,
execute directly and validate through the gate; do not loop back with
a draft when the work is yours to do.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
The existing FULL DELEGATION manual-QA rule said 'use it yourself' but
left the choice of tool implicit. Make it explicit and non-optional, so
the agent cannot satisfy the gate by reading the source instead of
running the artifact.
Surface → tool mapping:
- TUI / CLI work → interactive_bash (tmux). Launch in real terminal,
send keystrokes, run happy path, try bad input, hit --help.
- Web / browser / UI work → playwright skill. Drive a real browser,
click elements, fill forms, watch console, screenshot if helpful.
- HTTP API / service work → curl or integration script against the
running service.
- Library / SDK work → minimal driver script that imports + executes.
- Other surfaces → ask how a real user would discover it works, then
do that.
Frame the gate as a contract violation when bypassed: reporting
'implementation complete' without using the matching tool is the same
failure pattern as deleting a failing test for a green build.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Existing <verification> required tests pass + lsp clean + build green, but
that is insufficient for end-to-end delegation. Tests cover known cases;
they do not cover whether the user-visible feature actually works.
Add a NON-NEGOTIABLE rule: when the user hands off end-to-end ("ulw",
"implement and finish", "do the whole thing", "make it work", "ship it"),
verification escalates to:
1. BUILD the actual artifact
2. USE IT YOURSELF as a real user would
3. VERIFY end-to-end behavior matches the spec
4. TASK NOT DONE until usage confirms it works
Reporting "implementation complete" without having USED the artifact is
explicitly framed as a contract violation. Defects discovered during this
QA pass are the agent's to fix in the same turn.
This complements the existing 'lsp_diagnostics catches type errors, not
logic bugs' line by giving full-delegation cases a sharper, named gate.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
YeonGyu-Kim