compactedSessions permanently blocked re-compaction after first success,
causing unbounded context growth (e.g. 500k on Kimi K2.5 with 256k limit).
- Clear compactedSessions flag on new message.updated so compaction can
re-trigger when context exceeds threshold again
- Use modelContextLimitsCache for model-specific context limits instead
of always falling back to 200k for non-Anthropic providers
Hashline edit tool and companion hooks now require explicit opt-in
via `"hashline_edit": true` in config. Previously enabled by default.
- tool-registry: hashline edit tool not registered unless opted in
- create-tool-guard-hooks: hashline-read-enhancer disabled by default
- Updated config schema comment and documentation
- Added TDD tests for default behavior
Previously decoded entire image buffer to read headers. Now slices base64
to 32KB prefix before decoding — sufficient for PNG/GIF/WebP/JPEG headers.
Dramatically reduces memory allocation for large images.
applyToolConfig() unconditionally set question permission based only on
OPENCODE_CLI_RUN_MODE, ignoring the question:deny already configured via
OPENCODE_CONFIG_CONTENT. This caused agents to hang in headless environments
(e.g. Maestro Auto Run) where the host sets question:deny but does not
know about the plugin-internal OPENCODE_CLI_RUN_MODE variable.
Read permission.question from OPENCODE_CONFIG_CONTENT and give it highest
priority: config deny > CLI run mode deny > default allow.
Ripgrep's --glob flag silently returns zero results when the search target
is an absolute path and the pattern contains directory prefixes (e.g.
'apps/backend/**/*.ts' with '/project'). This is a known ripgrep behavior
where glob matching fails against paths rooted at absolute arguments.
Fix by running ripgrep with cwd set to the search path and '.' as the
search target, matching how the find backend already operates. Ripgrep
then sees relative paths internally, so directory-prefixed globs match
correctly. Output paths are resolved back to absolute via resolve().
Intercepts Read tool output with image attachments and resizes to comply with Anthropic API limits (≤1568px long edge, ≤5MB). Only activates for Anthropic provider sessions and appends resize metadata (original/new resolution, token count) to tool output.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Non-Claude models skip planning and under-parallelize. Two new sections
injected only when model is not Claude:
- Plan Agent Dependency: multi-step tasks MUST consult Plan Agent first,
use session_id for follow-ups, ask aggressively when ambiguous
- Deep Parallel Delegation (rewrite): explicit '4 units = 4 agents'
pattern, each with clear GOAL + success criteria, all run_in_background
Sisyphus prompt instructed 'your next action is background_output' which
caused agents to repeatedly poll running tasks instead of ending their
response and waiting for the system notification.
- Replace 'STOP all other output' with 'end your response' (actionable)
- Add system-reminder notification mechanism explanation
- Add explicit 'Do NOT poll' prohibition
- Reduce background_cancel(all=true) mentions from 5x to 1x (Hard Blocks)
- Reduce Oracle collect obligation from 4x to 2x
- Remove motivational fluff ('blind spots', 'normal and expected')
Net: -2 lines, clearer mechanism, eliminates polling loop root cause.
YeonGyu-Kim