TimemAi v0.8

TimemAi v0.8

Added

• Tail-aware KV-cache planning for growing prompt deltas.
• Replay tests that simulate provider-side cache matching and guard against low-hit-rate cache strategies.
• CI coverage for KV-cache replay gates and openai-compatible cache marker generation.

Changed

• Refined the model-facing response envelope wording and regenerated the expanded static prompt snapshot.
• Documented the KV-cache tail planning algorithm and replay evidence.

Fixed

• Stabilized Thought / Action panel rendering when ANSI color sequences are present, preventing visible-width miscalculation during long command/status redraws.
• Stabilized the cross-process mem guard serialization test on macOS CI with an explicit lock handoff.

Validation

• Local production gate passed: scripts/ci.sh.
• GitHub Actions CI passed on macos-latest and ubuntu-latest for main run 28648210606.

TimemAi v0.7.1

Patch release focused on prompt protocol clarity.

Changes:

• Removed Markdown fenced code blocks from the model-facing static prompt, response schema summary, and generated tool action examples.
• Updated the expanded static prompt snapshot so the audited prompt text matches runtime behavior.
• Added regression checks to prevent reintroducing prompt fences.

Quality gate:

• scripts/ci.sh passed, including Rust workspace tests, edge regression, release build, and real TTY smoke.
• scripts/sensitive_scan.sh --current passed.
• scripts/sensitive_scan.sh --history passed.

This release intentionally does not include the kvc_optimize branch changes.

TimemAi v0.7

TimemAi v0.7

This release focuses on making the agent runtime more modular, inspectable, and resilient before broader use.

Highlights

• Reworked the static prompt source to Markdown with explicit injection points for response schema, tool catalog, and skill headers.
• Added manifest-backed capability metadata so one tool description source drives both prompt rendering and executor-side validation.
• Moved tool-specific validation out of the top-level parser and into the capability/executor boundary.
• Added self_tool for Timem self-inspection, including non-secret runtime env, memory/audit paths, and about/version metadata.
• Added concise model-facing tool documentation instead of verbose schema dumps.
• Added focused scenario replay coverage for coding, memory QA, self QA/env update, file writing, KV-cache planning, and profiler cached-token accounting.

Stability

• Provider/network failures now retry up to five times with user-visible retry status.
• Protocol repair now returns focused repair context around the malformed model output.
• Repair overhead is visible as ⇌N (⚠M) in status lines.
• Premature finished responses with evidence-gathering actions are downgraded to working, so runtime executes the actions first.
• API audit is now structured as api_audit.json while legacy JSONL chat audit remains readable.

Quality Gate

• scripts/ci.sh: passed
• cargo test --workspace: passed
• scripts/edge_regression.sh: passed
• scripts/sensitive_scan.sh --current: passed
• scripts/sensitive_scan.sh --history: passed
• cargo build -p timem_shell --release: passed

TimemAi v0.6

TimemAi v0.6

This release hardens TimemAi as a local shell agent with stronger runtime
protocols, memory management, terminal UI behavior, and release gates.

Highlights

• Unified model-facing memory/context protocol under memmgr, covering durable
  memory, raw chat history, scratch memory, and prompt-context shrink.
• Replaced finality handling with report_job_progress plus continue, so
  progress can appear in the Thought/Action panel while long jobs continue.
• Added guarded finalization: continue:false plus runtime-controlled expect
  checks can finish without an unnecessary extra model round.
• Added compact thinking/final status layouts with model round count, context
  utilization, token totals, latest request deltas, and cache-hit counts.
• Increased default maximum agent interaction rounds from 20 to 50.

Reliability And UX

• Observation panel wraps long intent/action lines and renders action details as
  tree child rows under user-facing intent.
• Model-private thought content is hidden from the user-facing observation UI.
• Prose-wrapped or fenced JSON model responses are parsed when they contain a
  valid response envelope.
• Paste recovery no longer flags untouched [ pasted N lines ] markers from
  stale preserved paste records, and Esc cancels the current input activity.
• Final response status uses concise ctx[N%] instead of mixing current-turn
  deltas into completed summaries.

Quality Gates

• Added GitHub Actions CI that mirrors local release gates.
• Release validation passed locally with scripts/ci.sh, including script
  checks, install logic, contract checks, sensitive scan, Rust format/tests,
  release build, real TTY smoke, and whitespace checks.

v0.5.2

v0.5.2 — 2026-06-30

Bug Fixes

• shell: 修复 reedline 粘贴时的终端重绘异常，提升多行内容粘贴的稳定性 (c0a3b9f)

Improvements

• shell: 优化编辑后粘贴（edited paste）的恢复流程，减少内容丢失或光标错位问题 (92ffd77)

Docs

• shell: 补充 Ctrl-C 与 Esc 取消行为的说明，明确两者在不同场景下的差异 (5fd88b9)
• 记录 v0.5.1 后续审计结论 (705042a)

Tests

• 新增功能覆盖审计门控（feature coverage audit gates），确保关键路径有测试保障 (94376bb)

TimemAi v0.5.1

TimemAi v0.5.1

Reliability and regression-fix release after v0.5.

Fixed

• Prevent forced-shrink loops when the input context window is intentionally reduced.
• Tighten token status rendering, including the compact context label format: [ctx 8.2K].
• Avoid showing latest output-token deltas in final response status.
• Avoid misleading zero-total token delta display while a request is in progress.

Improved

• Add typed scratch context offload support so prompt delta/slice content can be moved into scratch by id.
• Clarify the forced shrink compaction protocol around scratch offload and prompt shrink.
• Strengthen repeated end-to-end shell edge regression coverage in CI.
• Document and enforce feature test coverage expectations.

Quality Gate

• Full scripts/ci.sh passed locally before release.
• Sensitive scan passed before release.
• Worktree was clean at tag creation.

Tag: v0.5.1
Commit: d39af91

TimemAi v0.5

TimemAi v0.5

v0.5 focuses on making the standalone Timem shell agent reliable enough for daily use and external release. The biggest change is the new terminal input layer, plus a stronger release gate so future changes are tested the same way users actually run the shell.

Highlights

More reliable terminal input

• Replaced the custom input handling with a Reedline-backed editor.
• Added Shift+Enter multiline input support.
• Improved mixed Chinese/English input, long wrapped lines, deletion, cursor movement, and Ctrl+C cancellation behavior.
• Multiline paste now collapses into a clear placeholder such as [ pasted 3 lines ], while preserving the original pasted content for submission.
• Edited paste placeholders are detected and can be restored before submission.

Better release quality gates

• Added scripts/ci.sh as the release gate for formatting, tests, release build, real TTY smoke, install-script checks, and sensitive-info scanning.
• Added a real pseudo-terminal smoke test that drives the compiled shell binary instead of only testing pure functions.
• Added sensitive-information scans for the current tree and git history.
• Added install-script logic tests so dependency/version checks are covered before publishing.

Documentation and runtime layout cleanup

• Updated architecture documentation for the shell runtime modules: input editor, observation panel, profiler, prompt cache, protocol adapter, and structured output handling.
• Fixed runtime data documentation to match the current layout:
  • data/<space>/audit/api_audit.jsonl
  • data/<space>/audit/action_audit.json
  • data/<space>/memory/
  • data/<space>/shell_history.txt
• Updated the v0.5 release audit with the Reedline input work and the new quality gates.

Upgrade

From an existing git clone:

git pull --ff-only
./install.sh

Then load your private env file and run Timem:

source /path/to/your/env
timem

To run the latest source directly without reinstalling:

cargo run -p timem_shell

Validation

The v0.5 tag points to:

1c59365036567f029ae3b1ff4f34e55c7eccfabb

Release validation completed on main:

scripts/ci.sh
scripts/sensitive_scan.sh --history

Both passed. The CI gate includes Rust tests, release build, real TTY smoke, install logic checks, and current-tree sensitive scanning.

Notes

• Windows is still not supported. macOS and Linux are the supported targets for this release.
• Private API keys should stay in a local env file and should not be committed.
• v0.5 is a source release; install from the git clone with ./install.sh.

v0.4

TimemAi v0.4

TimemAi is a lightweight local agent for everyday work. It combines an LLM conversation loop, guarded Bash command execution, and local multidimensional structured memory so it can inspect your working environment, help with files and commands, and remember useful context across sessions.

• Add README banner and project positioning: Bash, Memory, Time-aware.
• Publish standalone Timem shell package with install/update/uninstall workflow.
• Use explicit env sourcing via env_template and document cargo run for latest local builds.
• Add gateway provider/API protocol/base URL configuration and startup config banner.
• Improve prompt delta/slice handling, shrink accounting, cache marker display, and audit paths.
• Harden LLM protocol parsing for escaped/variant model outputs.
• Harden terminal approval/input recovery so paste, arrows, Ctrl-C, and CR control sequences do not corrupt shell input.
• Clean release examples so local machine paths, env files, audit logs, and data files are not tracked.

Validation: cargo test -p agent_core and cargo test -p timem_shell passed before this v0.4 update.