Add Maki as a supported CCGram provider

[UN-9BOT]

Problem

CCGram currently does not support Maki as a first-class provider. Setting CCGRAM_PROVIDER=maki falls back to Claude because maki is not registered in the provider registry.

https://github.com/tontinton/maki

This means users who run Maki inside tmux or herdr cannot:

• select Maki from the provider picker
• launch Maki sessions from CCGram
• resume or continue Maki sessions from CCGram
• have Maki transcripts parsed and relayed correctly
• auto-detect Maki panes/windows as a known provider

Why this matters

Maki is an actively used terminal coding agent with:

• resumable sessions
• persisted JSONL session storage
• a Claude-compatible stream-json mode
• built-in slash commands and task/subagent workflows

Without provider support, CCGram cannot manage Maki sessions in the same way it supports Claude, Codex, Gemini, Pi, and Antigravity.

Proposed change

Add initial maki provider support to CCGram with:

• provider registry wiring
• launch command resolution (maki, CCGRAM_MAKI_COMMAND)
• resume/continue support
• Maki session discovery from state dirs
• transcript parsing for Maki session JSONL files
• runtime/provider detection from command/path/transcript path
• provider picker and toolbar integration
• tests for the new behavior

Environment

• CCGram version: local dev build from current repository branch
• Multiplexer: herdr

[alexei-led]

Keeping this open for justification before investing in a first-class provider. Please add:

• a link to the upstream Maki project and the concrete ccgram use case;
• why first-class integration is needed instead of the generic shell/session path;
• expected users and maintenance commitment;
• supported OSes and authoritative state/session/transcript locations;
• the minimal acceptance criteria for session discovery, streaming, resume, permissions, and errors.

This is the first request we have received for Maki, so we need evidence that the integration and ongoing compatibility cost are justified.

[UN-9BOT]

Thanks for the detailed rejection notes on #159 — they were the right
list. I went back through each item and pushed a new commit on
feat/maki-provider that addresses them.

What changed since #159 was closed:

1. Reliable state-discovery contract (POSIX-only, not cross-platform)

• New module src/ccgram/providers/_maki_paths.py is the single source
  of truth for where Maki state lives.
• Resolution order, documented in the module docstring:
  1. MAKI_STATE_DIR env override (for CI / sandbox)
  2. $XDG_STATE_HOME/maki
  3. ~/.local/state/maki
  4. ~/.maki (legacy)
• Only directories that actually exist are returned.
• Windows is explicitly out of scope and documented as such —
  ccgram is a POSIX-only project, and the docstring spells that out so
  it isn't a hidden assumption.

2. Bounded rendering policy for tool output

• New helpers render_tool_result / render_diff_payload in
  src/ccgram/providers/maki.py with module-level constants
  (TOOL_RESULT_MAX_CHARS = 4000, TOOL_RESULT_MAX_LINES = 80,
  TOOL_RESULT_DIFF_MAX_LINES = 40, TOOL_RESULT_DIFF_HEADER = "[diff]").
• Raw out payload text no longer reaches Telegram unbounded; it's
  either returned as-is (within limits), wrapped in
  format_expandable_quote (over the line cap), or truncated with an
  explicit … [truncated N char(s)] marker (over the char cap).
• Diff payloads get a [diff] header and are bounded separately.

3. Transcript summarization no longer treats a tool result as a user prompt

• Session title candidate is now strictly
  t == "msg" AND d.role == "user" AND non-empty text.
• _user_title_text rejects tool_result / tool_use entries.
• Covered by contract test test_summary_skips_tool_result_only_transcripts.

4. Unrelated Edge TTS fix separated

• Lifted out of the Maki branch into its own branch + PR.
• fix/edge-tts-test-import → PR test: avoid importing optional edge-tts at test-collection time #167 against main.

5. PR checks are green

• make lint → All checks passed!
• make test → 6197 passed, 30 skipped
• Maki-specific: 30 tests pass.

6. Focused contract + integration tests

• New file tests/ccgram/providers/test_maki_contracts.py covers:
  • locator contract (empty / override / XDG / legacy / nonexistent)
  • rendering policy (short / empty / char-bound / multiline / diff)
  • summary candidates (user text only, fallback to id)
  • launch args (incl. invalid resume_id rejection)
  • capability descriptor (hookless, conservative flags)

A few things I deliberately kept out of scope (and called out in the
PR body):

• Windows support — ccgram is POSIX-only.
• ccgram hook --install --provider maki — Maki has no equivalent
  hook surface; the provider stays hookless.
• Status snapshot / task tracking refinements — current capability
  values are intentionally conservative; happy to revisit in a
  follow-up once the basics land.

The branch feat/maki-provider on my fork now points at the new
implementation (54ebe11), with the previous commit (57407cd)
preserved in history. Want me to open a fresh PR off this branch (it
should reference this issue and not #159), or are there additional
requirements you'd like pinned down here first?