Repository navigation
v1.0.168
[1.0.168] - 2026-11-09
Spec system repair and documentation release.
Ten commits since the 1.0.165 documentation pass, of which only the spec
repair pair (6953c57c, 9ed32bd8) changes production code after 1.0.167.
The rest are the 1.0.166 research/teams/agents release, its review follow-up,
the 1.0.167 CI fix and the documentation passes. Each commit is listed below
with its hash.
Spec system repair (6953c57c, 9ed32bd8)
A review of the whole spec system - ragent-specs, the spec_* agent tools and
the TUI /spec handlers - repaired the SPEC.md/PLAN.md round-trips and a
set of /spec command defects, and made task completion times durable.
Fixed - persistence (ragent-specs)
- The PLAN.md task table parser no longer keeps empty cells, so a blank
Requirement cell no longer shifts every later column of that row. - The em/en dash placeholders the task-row formatter writes now read back as
"none". Before, every task update gave an unlinked task a requirement and a
dependency literally named-. - A task-status update rewrites the matching rows in place instead of deleting
everything between the task table and the next heading, and the trailing
newline is preserved. - The frontmatter audit trail is read back, so history survives each save
instead of being regenerated empty. - A status change keeps the frontmatter keys it does not own (
research,
custom keys). - Frontmatter closes only on a line that is exactly
---, so a---inside a
value no longer truncates the block. - Multi-line
research:lists are parsed. - One unreadable spec no longer aborts discovery for every other spec.
/spec implhonourssdd.phase_minus_one_gateswhen moving a spec to
in_progress.- A new
implemented -> in_progresstransition reopens an implemented spec for
further work. - Task completion times are persisted in the PLAN.md Status cell as RFC 3339
UTC, e.g.completed (2026-11-08T10:12:00Z); no extra column is added, so 6-
and 7-column task tables keep parsing. - YAML comment lines no longer count as top-level frontmatter keys, so a comment
inside areviewers/research/auditlist does not truncate it and the
writer no longer copies orphaned list items (io::is_frontmatter_key_line). - The PLAN.md task rewrite now touches exactly what the plan parser reads as
tasks: the first task header in the Tasks section (a legacy
| Task ID | Title |header is normalised to seven columns too) and data rows
with at least six cells. An unrelated table that merely mentions a task ID,
such as a three-column review table, is left untouched.
Fixed - /spec commands (TUI)
/spec statusreported "from X to X" because it read the status before the
transition; it now reports the transition it performed./spec listaccepts the documented--status <s>form and rejects an
unknown status or flag instead of listing everything.- A flag followed by another option (
--prefix --archived) or an empty
--status=is a missing-value error instead of swallowing the next token. - Deleting the active spec deactivates it, and
/spec deactivateinvalidates
the cached spec section of the system prompt. /spec implguidance and its completion message reflect the real status, and
a failed write toimplementedis reported as an error (SPEC.md keeps its old
status) instead of being read back from the mutated in-memory status as
success.- Status lines no longer carry a doubled
spec: spec:prefix, and an invalid
spec ID is explained in the message window. - The
/spec status,/spec taskand/spec activatehandlers are
rustfmt-formatted again.
Changed - documentation
SPEC.mdredraws the spec lifecycle diagram to match
SpecManager::allowed_transitionsand adds section 10.10e (task completion
timestamps).
Tests
crates/ragent-specs/tests/test_spec_roundtrip.rs(round-trip and
frontmatter cases),crates/ragent-tui/tests/test_slash_spec_deep.rs,
crates/ragent-tui/tests/test_event_swarm_research_deep.rs, plus additions
totest_manager,test_spec_impland the inlineplan_parser,
impl_runnerandmanagermodules.
161a5f47 - merge pull request #11 (fix/spec-review)
Merge of the spec repair pair above; no additional change.
1e1d33a8 - Version: 1.0.167
CI fix release; no production code change.
test_mcp_acp_more::acp_cancel_flag_aborts_a_stalled_turnno longer fails
intermittently withleft: Exited, right: Protocol. The four ACP tests in
that file shared one scratch script path (more_agent.py), and
write_scripttruncates before writing, so undercargo nextest- one
process per test - a sibling could handpython3a half-written script. The
interpreter then died with aSyntaxError, which the transport reports as
AcpErrorKind::Exitedinstead of theProtocolcause the test asserts. The
script path is now per-behaviour (more_agent_<behavior>.py).
a5fb4289 - Version: 1.0.166
Research, teams and agents release.
/researchrepaired end to end:continueresumes instead of printing
"Resuming",showprints each source's own kind and a 1-based id through one
sharedshow_source_rowshelper, a second run for an in-progress item is
refused instead of mixing two runs' progress, a failed run no longer strands
the status bar on[wait], an unknown--flagor a value flag given no value
is rejected instead of being folded into the item name, undated progress text
is no longer erased by the next refresh, andragent.jsonload failures are
reported.- Teams: the runtime is bound in every entry point (TUI,
ragent run,
ragent serve, ACP) rather than only the TUI and is rebindable
(TeamManagerSlot/clear_if); every team tool exceptteam_create
resolves its caller and refuses lead-only actions; the mailbox poll loop peeks
instead of marking messages read; an idle teammate is woken on mail,
auto-claimed work, resume and shutdown; everyconfig.jsonread-modify-write
goes throughTeamStore::updateunder one lock. - Agents:
wait_agentswaits on running background tasks again (regressed in
1.0.153), and oneTaskStatus::is_terminal()predicate decides "finished"
on every path.
b732a3a5 - docs: sub-agent lifecycle, team runtime binding and v1.0.165 stats
Documented the ten commits since v1.0.165 across QUICKSTART.md,
TUI-QUICKSTART.md, README.md, SPEC.md and STATS.md, updated
docs/howtos/tools/sub-agents.md (wait seeding, one definition of "finished")
and docs/howtos/teams.md (6.6.2 caller resolution, lead-only tools, runtime
binding, rebindable TeamManagerSlot, atomic TeamStore::update), and
regenerated 130 tracked howto PDFs.
e02d2678 - merge pull request #10 (fix/subagent-lifecycle)
Merge of 77826b3b, fe4f0b46 and ad52c257; no additional change.
ad52c257 - fix(agents): address review - drain flag, suspend races, event payload
Review follow-up on the sub-agent lifecycle work.
spawn_syncreturns the task snapshot taken under the completion guard, so a
concurrent drain reaping the newly reported entry cannot turn a successful run
into "vanished after completion".suspend_task/resume_taskstore the park flag under the task guard, so an
interleaved suspend/resume cannot leave a running task parked.has_pending_backgroundis no longer cleared while any background task is
live or another parent has a pending result, and each background completion
re-arms it.SubagentCompletecarries the terminal status andoutput_file, so the TUI
no longer needs a registry lookup that could race with reaping.- Reconciliation increments the active-agents gauge when it re-adds a row, and
note_subagent_cancelledrecordsCancelled, notTerminating. - Two older tests that asserted the buggy behaviour were updated.
fe4f0b46 - docs(changelog): sub-agent lifecycle fixes
CHANGELOG-only commit recording the 77826b3b fixes.
77826b3b - fix(agents): sub-agent lifecycle, wait_agents and Agents panel bugs
- Task registry: adds
TaskStatus::is_terminal();drain_completedno longer
injects and reaps suspended/terminating tasks;kill_taskrefuses a
timed_outtask;suspend_task/resume_taskcheck and set status under one
guard;cancel_allalso cancels suspended tasks; finished foreground
(spawn_sync) entries are marked reported and reaped by the next drain. - Tools:
wait_agentswithtask_idsomitted waits for running background
tasks again;new_agentforeground metadata reports the registry status
instead of a hard-codedcompleted. - TUI Agents panel: history records failed/timed-out runs with their real status
and the durableoutput_file; a plain kill keeps the rowterminatingwhile
a force-kill records a failed history row and decrements the active-agents
gauge; reconciliation keeps terminating rows and retires finished ones into
history; a failed completion logs[err] Task failed. - Tests:
test_subagent_lifecycle_reviewinragent-agent(8 cases) and
ragent-tui(5 cases).