The absolute 16,384-token floor plus the carried summary and output
reserves exceeds windows below ~58k outright, and overflow recovery then
bailed at the very floor that caused the rejection, leaving compaction
unusable on small-context models. Scale the floor to window/8 (min 1k)
and use the same floor in overflow recovery.
A session that crossed a provider boundary compacted 90 times in three days
without ever succeeding: every attempt asked the summarizer to read the whole
re-expanded span in one call (2.33M tokens on 08-15, 3.03M by 08-17, against a
1M cap), and every rejection was retried ten times.
Three independent defects:
1. `generateSummary` serialized the entire span into one prompt with no budget
check. It now plans windows that fit the summarizer's context and folds them
with the update prompt that iterative compaction already uses, so a stranded
boundary is recovered instead of rejected. A provider that rejects a window
the catalog said would fit (claude-sonnet-4-5 advertises 1M but is
beta-gated to 200k on OAuth credentials) halves what was actually sent and
re-plans, because only the rejection knows the real cap.
2. `TRANSIENT_TRANSPORT_PATTERN` matched bare status codes, so the random id in
the `raw-http-request=.../1787022540720-3o503gxo48bvb.json` pointer omp
appends to its own errors classified a deterministic 400 as a transient 503.
Statuses are now word-boundaried, matching AUTH_FAILURE_PATTERN.
3. Neither retry layer vetoed ContextOverflow, so one failure became up to 30
identical calls (10 outer x 3 oneshot). A oneshot replays a fixed prompt, so
an input that does not fit never fits; both layers now fail fast to the next
candidate.
The boundary scan that decides which compaction entry a model can actually read
is extracted as `findReadableCompactionIndex`, since the fold and
`prepareCompaction` both need it.
Verified by replaying the session that failed: 7,096 messages summarize in 3
calls with a largest prompt of 773,705 tokens under the real 1M cap, and in 15
calls with a largest prompt of 196,148 tokens under a simulated 200k cap.
Retry: widened agent dequeue-hook deadline budgets from 25ms to 1s — the run loop checks the deadline before invoking dequeue hooks, so a cold or CPU-starved mock roundtrip expired the deadline first and the hooks never ran (deterministic failure in isolation, flaky under CI parallel load).
prepareCompaction walked the branch from the last compaction and ignored reset_boundary markers, so /compact (and auto-compaction) resurrected pre-/clear turns into the summary even though buildSessionContext already starts the model context after the boundary.
Model reset_boundary as a first-class agent-core session entry and start the summarization window after the latest boundary, dropping the superseded pre-reset compaction summary. A boundary before the last compaction stays superseded by it.
Fixes#8718
- Retained tool-search server calls and opaque results in signed assistant history across direct streams, gateways, and custom-endpoint projection.
- Added replay regressions for interleaved thinking and client tool continuations.
Fixes#8559
- Replaced time-based sleeps and polling loops with event-driven promise resolvers and fake timers across agent and tool tests.
- Migrated test suites to share in-memory auth storage and fixtures using lifecycle hooks.
- Updated catalog model definitions, metadata, and configurations.
- Remove redundant definedness, null, and type checks across test suites in multiple packages.
- Clean up unused assertions, metadata tests, and obsolete test cases.
- Add good versus bad test filter guidelines and requirements to project documentation.
Each is a single, side-effect-free completion whose result is parsed after
it resolves, so one transient provider failure previously aborted the
whole operation - for /compact that left the user's context full.
Adds SummaryOptions.oneshotRetry, because both compaction paths call the
same generateSummary and the policy therefore cannot be a constant inside
it. Manual /compact has no outer loop and gets retry by default;
auto-compaction passes false because session-maintenance already retries
the whole attempt, and nesting would multiply the budget (10 outer x 3
inner) while stacking each outer wait on an inner backoff.
instrumentedCompleteSimple is the single funnel for every oneshot LLM
call in the agent, so retry belongs here rather than in a try/catch above
each caller - the failure arrives as a resolved AssistantMessage.
Opt-in rather than default-on: oneshotKind is free-form and callers may
pass arbitrary ctx.tools, so the funnel cannot itself prove a request is
replay-safe.
Response headers are captured per attempt and cleared between attempts,
so a stale retry-after can never be reused for a later failure.
- Refactored and condensed numerous system prompts, agent instructions, and tool documentation files across packages.
- Streamlined workflow rules, formatting constraints, and execution guidelines for improved clarity and brevity.
- Updated discovery rules, recommendation criteria, and syntax standards in prompt templates.
RESCUE_SHAKE_CONFIG spreads AGGRESSIVE_SHAKE_CONFIG, so the new 4k manual
tail leaked into dead-end recovery and could block eliding the very
result that caused the dead end. Override protectTokens back to 0 in the
rescue preset and add the coding-agent changelog entry for the manual
/shake behavior change.
Addresses review on #8067.
Cursor bash/grep frames wrote optional kwargs as present-undefined, and
tools.format gemini projectors dropped kCursorExecResolved so settled
calls ran twice.
Co-authored-by: Cursor <cursoragent@cursor.com>
Manual /shake used protectTokens: 0, stripping every eligible tool result
including the ones the agent is still working from. Keep a small 4k-token
recent window (matching the automatic shake mechanism, at a quarter of its
16k budget) so the full escape hatch stays aggressive without destroying
the live tail. Two matcher-focused tests that implicitly relied on the
zero window now pin protectTokens: 0 explicitly.
Fixes#7776
When an xd:// device is dispatched through the write tool, the outer
approval gate now consults tools.approval.<deviceName> before falling
back to tools.approval.write. This lets users scope allow/deny/prompt
to a single device mount without changing the blanket write tool policy.
The write tool's approval function returns { tier, policyKey: deviceName }
for xd:// device dispatches. resolveApproval uses the policyKey to look
up the user override on the device name, falling back to the invoking
tool's own policy when the device has none configured.
Adds:
- ToolApprovalDecision.policyKey field (optional, additive)
- policyKey-aware lookup in resolveApproval and requiresApproval
- Updated error messages naming the correct config key
- Unit tests for policyKey resolution and WriteTool integration
Fixescan1357/oh-my-pi#7923
can1357