- fix(compaction): tokensBefore undefined crash on reload
compaction-orchestrator now falls back to preparation.totalTokens when
extension returns tokensBefore: undefined; compaction-summary-message
guards with ?? 0 defensively
- feat(exec): inline truncation notice in sf_exec digest
appends [stdout truncated — read full output: <path>] when
stdout_truncated=true so agent knows to use sf_exec_search
- feat(exec): wire onUpdate progress for sf_exec
calls onUpdate before execution starts with status/command so TUI
shows live feedback during long-running commands
- feat(security): prompt injection defense for external content
new sanitize-external-content.js utility: strips HTML comments,
detects 15 injection patterns (instruction override, role reassignment,
fake system messages, encoded payloads); wired into exec-tool digest
- feat(tools): sf_session_todo tool (persisted cross-compaction)
add/check/list ops; persists to .sf/session_todo.json; pending todos
injected into compaction summary block for context continuity
- feat(hooks): shell hooks surface (.sf/hooks/pre-tool/*.sh, post-tool/*.sh)
pre-tool hooks block tool execution (exit≠0 = block with stdout reason)
post-tool hooks fire-and-forget; JSON context piped to stdin; 5s timeout
- fix(db): WAL autocheckpoint disabled to prevent corruption
PRAGMA wal_autocheckpoint=0 in initSchema(); explicit checkpointWal()
after successful finalize verification — the only safe checkpoint point
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When the autonomous solver fails to produce a checkpoint and enters the
repair loop, subsequent retries previously called newSession() each time,
wiping the conversation history. The agent restarted cold with no memory
of what it had tried, what tools it had called, or why it failed — making
meaningful repair nearly impossible.
This change adds a keepSession option to runUnit(). When true, the
newSession() call and session-switch guard logic are skipped; the repair
prompt is sent as a follow-up in the existing conversation. The agent can
now see its prior tool calls, file reads, and failure context when deciding
how to fix the issue.
Policy:
- First attempt at each unit: keepSession=false (clean session, correct
for independent slice boundaries — system prompt carries project state)
- Repair retries within the same unit: keepSession=true (agent carries
full context of what it already tried)
- Next unit after success/failure: keepSession=false (clean boundary)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When the selected model's provider is not request-ready:
1. Pre-flight check before runUnit: find any ready provider, switch to it
and continue. Only stop if no ready provider exists.
2. Post-runUnit cancelled handler: same logic — reselect + return 'continue'
instead of silently breaking.
3. Both paths now emit a visible ctx.ui.notify so the user can see what
happened ('provider X not ready — retrying with Y/model').
Previously: cancelled instantly, all 4 repair attempts also cancelled,
paused with misleading solver-missing-checkpoint and no user notification.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When runUnit() returns status='cancelled' (provider not ready, session
failed, timeout), there is no checkpoint to repair. Previously the code
called assessAutonomousSolverTurn() which saw no checkpoint and entered
the 4-attempt repair loop — all of which also cancelled instantly,
burning retries before pausing with a misleading solver-missing-checkpoint
reason instead of surfacing the real provider/session error.
Now: cancelled result short-circuits to { action: 'none' }, skipping the
repair loop and falling through to the existing cancelled handler which
correctly surfaces provider-not-ready, timeout, and session-failed errors.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
v1 no longer exists — the suffix is just noise. Update all import sites
and rename the test file to match.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
