v0.2.28

• Redesign the terminal UI around a compact dark/light semantic palette, with
  denser conversation cards, sidebar sections, footer controls, status surfaces,
  management panels, a centered session title, and a responsive rounded welcome
  card whose shortcut hint remains legible in both themes and narrow terminals.
• Add comprehensive mouse interaction: scroll the surface under the pointer,
  resize the sidebar, drag the conversation scrollbar, select and copy wrapped
  transcript text, activate code-copy controls and links, operate list panels,
  and expand any foldable conversation card from its full rendered area.
• Add a modal inline-image viewer with zoom, pan, reset, and protocol-aware
  caching. Kitty terminals upload one bounded PNG stream and update one cropped
  native placement while zooming or panning; pointer events and image work are
  coalesced so they do not starve keyboard input or model events.
• Render supported Mermaid flowcharts and mind maps directly in the terminal,
  with readable fallbacks for layouts that do not route reliably in character
  cells. Convert common inline, display, and fenced LaTeX-like math notation to
  terminal-friendly Unicode symbols and linear fractions.
• Improve command and background-task lifecycle handling around timeout,
  cancellation, process-tree cleanup, interactive terminals, and foreground
  commands moved into the task manager, avoiding stale registrations and races
  that could terminate an already-promoted task.
• Make active feedback calmer and cheaper: coalesce TUI pointer motion, preserve
  the 120 ms hover transition, animate complete pending tool and tool-group
  titles in one shared phase, and limit status-bar breathing to the activity icon
  so labels and elapsed time remain stable.
• Make session filesystem approvals operation-scoped and explicit. Existing
  directory approvals cover descendants, file approvals remain exact, stale
  sandbox approvals cannot tighten a newer mode, and approved paths are checked
  again before storage and child-policy propagation. Sandboxed child writes
  require overlapping recursive read and write approvals rather than silently
  widening write-only access.
• Tolerate brief cross-process memory-lock handoffs without failing ordinary
  recall and management operations, while keeping long-running Dream lock
  contention bounded and explicit.
• Expand regression coverage across pointer routing, selection, image protocols,
  process ownership, security grants, theme resolution, tool animation, memory
  locking, and responsive welcome rendering; refresh the user and architecture
  documentation.

Upgrade note: request_permission filesystem elevation now accepts only read
and write; executable access remains configurable through security policy but
cannot be granted interactively. An explicit session approval overrides a
matching configured deny for its approved operation and scope. Exact-file grants
apply to built-in path tools but cannot be represented safely in the native
subprocess sandbox, and filesystem grants do not constrain commands when process
sandboxing is off.

Full changelog: v0.2.27...v0.2.28

v0.2.27

• Give headless programmer run and programmer init stable exit codes:
  0 success, 1 runtime error, 2 usage error, 3 --check failed, 4 timeout,
  5 --max-steps reached, 10 input required, 130 interrupted.
• Add per-run security overrides that are never written back to the
  configuration: --security-profile, --sandbox restricted|network|off,
  --allow-read, --allow-write, --deny-read, --deny-env and
  --no-protect-file-changes. They are also accepted by diagnostics and the
  MCP server commands.
• Add run options: --config (loaded read-only), --append-prompt and
  --append-prompt-file, --image, --vision/--no-vision, --no-memory,
  --mcp/--no-mcp, --allowed-tools/--disallowed-tools,
  --coauthor/--no-coauthor and --soul-file. JSON and JSONL results now
  include stop_reason, work_mode, security, steps, tool_calls,
  duration_ms, session_id and compactions; JSONL adds start,
  compaction and input_required events. schema_version stays 1.
• Save and continue headless sessions with --save-session,
  --resume <UUID>, --continue, --fork and --no-save. Headless runs
  hold the same session lock as the TUI, so one session cannot be driven
  twice at once.
• Compact context during long headless runs, with --compact-tokens,
  --mandatory-compact-tokens and --no-compact overrides. The TUI's
  compaction logic moved to a shared module without behavior changes.
• Suspend instead of failing when a run needs a person: manual approvals,
  classifier Ask verdicts, ask_user, sandbox permission requests and plan
  review save the session and exit with code 10. The result carries a
  pending object with the question, options and a ready-to-copy
  resume_command; continue with --resume <UUID> --answer <ANSWER> (or
  --answer-file, optionally checked with --answer-id) from the original
  workspace. --work-mode manual suspends automatically; other modes need
  --suspend.
• Add unit and binary regressions for exit codes, overrides that leave the
  configuration untouched, sessions, compaction and suspension; update the
  headless documentation.

Upgrade note: scripts that treated any non-zero exit as one failure keep
working, but usage errors now exit 2, timeouts 4 and step limits 5 instead
of 1. A headless --work-mode manual run now exits 10 at the first approval
instead of denying the tool call.

Full changelog: v0.2.26...v0.2.27

v0.2.26

• Connect the configured mcp_servers in headless programmer run and
  programmer init, so the agent and its sub-agents get the same
  mcp__<server>__<tool> tools as the TUI. Previously headless runs started no
  MCP servers. Servers start after the work mode and classifier are validated;
  one that fails is skipped with a warning on stderr and the run continues.
• Connect MCP servers concurrently in both the TUI and headless runs. A slow or
  hung server now costs only its own handshake timeout instead of delaying every
  server configured after it. Startup errors keep configuration order, while
  connection status updates arrive in completion order.
• Share turn assembly between the TUI, headless runs, and sub-agents: one module
  installs the common tool providers, resolves the approval policy once for a
  parent and its children, and applies the same memory-model gate everywhere.
• Add regressions for shared assembly, the memory and classifier gates, headless
  MCP startup with a failing server, and concurrent startup past a hung server;
  update usage and architecture documentation.

Upgrade note: headless runs with mcp_servers configured now spend time
starting those servers before the first turn, and that time counts against
--timeout.

Full changelog: v0.2.25...v0.2.26

v0.2.25

• Fix compaction during long, tool-heavy turns: mandatory safe-point compaction
  includes completed work from the active turn, keeps tool calls paired with
  their outputs, and advances the summary boundary as new history accumulates.
• Stop repeatedly compacting summary-only history and correct full-history token
  estimates so differences between chat and compact providers do not invent a
  retained context tail. Refresh usage at safe points and start the soft-limit
  cooldown after successful mandatory compaction.
• Make background compaction and its provider retry cancellable with Esc. Ignore
  late results and pause queued input after cancellation instead of immediately
  restarting manual or automatic compaction.
• Improve Dream planning diagnostics with categorized HTTP failures and bounded
  cause chains, removing request URLs from transport errors to avoid exposing
  URL credentials. Preserve pending input when planning fails.
• Add release safeguards: exclude local drafts and runtime files, require green
  CI on the exact release commit, pin builds to the validated SHA, and verify
  complete release notes and all ten uploaded archives before public publication.
• Expand regression coverage for long-turn compaction, cancellation, historical
  summary placement, Dream failures, and release policy; update documentation.

Full changelog: v0.2.24...v0.2.25

v0.2.24

programmer v0.2.24

Includes the ownership and argument-generation changes listed below for the
withheld v0.2.23 draft, plus:

• Fix Windows PTY termination by owning a duplicate process handle and correctly
  interpreting Win32 termination results; tolerate only confirmed process exit,
  while preserving genuine termination errors.
• Add Windows regressions for independent handle ownership, repeated termination,
  already-exited children, and invalid handles. Keep the PTY shutdown regression.
• Allow the large (8 MiB) synthetic SSE regression sufficient time on loaded hosts.

Full changelog: v0.2.22...v0.2.24

---

programmer v0.2.23 (withheld draft)

• Separate UI, execution-loop, and session ownership; preserve interrupted
  responses and isolate task, agent, todo, and checkpoint state across sessions.
• Seal, cancel, and join session work before replacement or shutdown; retain
  retryable state on close/save failures without stranding queued input.
• Keep in-flight MCP connection snapshots valid during reload and reject stale
  provider refresh results, including outdated errors and notification counts.
• Recover delegated task text from durable identity-checked evidence and return
  Started status to the source without inferring completion or persisting consent.
• Return full Todo IDs and distinguish cancelled approvals from classifier denials.
• Stop runaway tool-argument generation before execution, return sanitized paired
  tool failures through the normal agent loop, and bound consecutive failures.
• Expand protocol, lifecycle, shutdown, persistence, and local MCP regressions;
  document actual TUI/process boundary acceptance and current ownership diagrams.

Full changelog: v0.2.22...v0.2.23

v0.2.22

programmer v0.2.22

• Add browsable Dream audits with source excerpts, inline before/after changes,
  conflict-checked whole-run rollback, and explicit interrupted-transaction recovery.
• Add a read-only session relationship graph and event timeline with persistent
  ancestor selection, padded rows, keyboard/mouse navigation, and narrow layouts.
• Keep Dream preview responsive and cancellable without consuming queued inputs;
  distinguish abandoned generation records from unconfirmed activity without
  rewriting audits, and show readable relative times in Dream status.
• Cache Dream detail layouts, render only visible content, and share conversation
  Markdown and code-block styling for source excerpts.
• Centralize ready-work dispatch so cancellation and compaction cannot strand
  queued input, while preserving draft, approval, and mandatory-compaction gates.
• Show mandatory compaction waits instead of stale Running tools status, group
  consecutive tools starting at two calls, and restore accepted input suggestions
  when the draft is deleted.
• Expand lifecycle, rendering, rollback, scheduling, and input regression coverage
  and update usage and architecture documentation.

Full changelog: v0.2.21...v0.2.22

v0.2.21

programmer v0.2.21

• Make operation ownership explicit and preserve queued requests until cancellation
  finishes, without bypassing drafts, approvals, or peer consent.
• Show the last observed execution stage while cancelling, and interrupt pending
  hooks and memory-association waits promptly.
• Publish explicit diagnostics results to the sidebar from the same snapshot,
  rejecting stale updates and retaining known findings after incomplete checks.
• Preserve session files on load failures and keep unsaved changes after save
  failures, without automatic idle retries.
• Isolate task managers per runtime and strengthen runner/tool boundaries with
  typed operation identifiers and narrow question handlers.
• Add lifecycle, persistence, task isolation, diagnostics, and cancellation
  regression coverage and update architecture documentation.

Full changelog: v0.2.20...v0.2.21

v0.2.20

What's Changed

• v0.2.20 by @huangdihd in #47

Full Changelog: v0.2.19...v0.2.20

v0.2.19

What's Changed

• v0.2.19 by @huangdihd in #46

Full Changelog: v0.2.18...v0.2.19

v0.2.18

programmer v0.2.18

• Remove keyword filtering from memory candidates and increase the default global
  candidate limit from 3 to 30. The memory model selects up to five relevant entries.
• Exclude memories already added or recalled in live context before applying
  candidate limits. Compacted-away records become eligible again; retained turns
  still count. Apply context tracking to interactive, headless, and child agents.
• Show unavailable or incomplete token accounting when responses omit usage or
  report all-zero counts, preserving the last positive context count.
• Allow exact filesystem permission requests while the process sandbox is off,
  including on Windows. Filesystem restrictions and user approval still apply.
• Render sequential edit_file batches as individual multiline diffs instead of
  displaying escaped JSON.

Full changelog: v0.2.17...v0.2.18