agent-context-budget

Name: agent-context-budget
Author: WenyuChiou

16stars4forks

How to install agent-context-budget?

wenyuchiou/agent-collab-skills/agent-context-budget

$npx -y skills add WenyuChiou/agent-collab-skills --skill agent-context-budget

Installs into the current project.

Summary

Use when multi-agent work risks context overflow, memory growth, noisy logs, oversized handoffs, cross-session continuation, or parallel Codex and Gemini execution.

Files · 1

View on GitHub

SKILL.md

1	# agent-context-budget
2
3	Context governor for multi-agent rounds. The core rule is simple:
4	keep `.coord/` as the canonical state, pass agents compact packets,
5	and never paste raw logs or unbounded memory into the main session.
6
7	## When to Use
8
9	Use this before or during:
10
11	- Large Codex + Gemini + Claude runs.
12	- Cross-session resumes where the prior conversation is too large.
13	- Any round with more than two delegate tasks.
14	- Any workflow where `.coord/memory.yml`, `.ai/_log_.txt`, or agent
15	summaries are starting to dominate the prompt.
16
17	Not for small one-agent edits. Use the direct delegate skill instead.
18
19	## Default Policy
20
21	If `.coord/plan.yml` lacks `context_policy`, add this block:
22
23	```yaml
24	context_policy:
25	main_session_token_budget: 3000
26	task_packet_token_budget: 6000
27	result_summary_word_budget: 250
28	memory_digest_token_budget: 1200
29	log_tail_lines_on_error: 50
30	raw_log_policy: path-only
31	agentmemory: optional
32
33	# W3 — $ cost gate per task (v0.2.2+)
34	# Optional. If set, agent-acceptance-gate flags any task whose
35	# codex/gemini delegate exceeded the cap. Estimated from token
36	# usage × provider price (Anthropic / OpenAI / Google rates).
37	# Set per-task in plan.yml tasks[].budget.max_cost_usd to override.
38	default_max_cost_usd: 0.50 # default $ ceiling per task
39	total_round_max_cost_usd: 5.00 # hard stop for the entire round
40	```
41
42	Why both `token` and `cost` budgets? Token gate prevents context
43	bloat (a session quality concern). Cost gate prevents runaway delegation
44	spend (a financial concern). They're orthogonal: a 2k-token delegate
45	call can cost $0.05 (Claude Haiku) or $0.50 (Claude Opus), so token
46	count alone doesn't bound dollars.
47
48	Per-task override example:
49
50	```yaml
51	# in plan.yml
52	tasks:
53	- id: T1
54	agent: codex
55	slug: simple-refactor
56	budget:
57	max_cost_usd: 0.10 # this task is mechanical, cap low
58	- id: T2
59	agent: codex
60	slug: complex-rewrite
61	budget:
62	max_cost_usd: 2.00 # this task needs frontier model, allow higher
63	```
64
65	## Workflow
66
67	1. Read `.coord/plan.yml` and `.coord/memory.yml` if present.
68	2. Write `.coord/context_<NNN>.md` with:
69	- round goal and success criteria
70	- task graph and write ownership
71	- per-agent context packets to include
72	- context intentionally excluded
73	- optional `agentmemory` recall queries, if available
74	3. Write or refresh `.coord/session_primer.md` with:
75	- current decisions
76	- open questions
77	- active round status
78	- recent artifacts by path
79	- no raw logs
80	4. Enforce result packets:
81	- delegate summaries are <= 250 words
82	- changed files are listed by path
83	- tests are listed by command + result
84	- risks are explicit and short
85	5. If a log is needed, include only the path. Read the last 50 lines
86	only when the corresponding `result.json` status is `error`.
87
88	## Optional agentmemory
89
90	`agentmemory` is a recall cache, not the source of truth.
91
92	- Query it only to enrich `.coord/session_primer.md`.
93	- Store only compact memory candidates: accepted decisions, resolved
94	open questions, artifact summaries, and final session outcomes.
95	- If it is unavailable, continue with `.coord/memory.yml` only.
96	- Never use vector recall as acceptance evidence. The acceptance gate
97	must read `.coord/plan.yml`, result files, reconciliation, and tests.
98
99	## Memory Promotion Rules
100
101	Promote only:
102
103	- Decisions with a one-sentence `what` and `why`.
104	- Open questions with a blocker and suggested next owner.
105	- Artifact pointers with one-line summaries.
106	- Agent session outcomes with status and result-summary path.
107
108	Do not promote:
109
110	- Raw logs.
111	- Full diffs.
112	- Source code.
113	- Long analysis.
114	- Secrets or credentials.
115
116	## Output
117
118	End with:
119
120	```text
121	[agent-context-budget]
122	Round: <N>
123	Policy: .coord/plan.yml context_policy
124	Context plan: .coord/context_<NNN>.md
125	Session primer: .coord/session_primer.md
126	Raw logs: path-only; failure tail max 50 lines
127	agentmemory: optional cache, not required
128	```
129
130	## Common Mistakes
131
132	- Pasting full logs into chat. Store the path; read bounded tail only
133	on failure.
134	- Treating memory as a transcript. Promote decisions, questions,
135	artifacts, and outcomes only.
136	- Giving Gemini only `.ai/` paths. Inline critical context because
137	`.ai/` may be gitignored.
138	- Using agentmemory as the gate. Gate with tests and `.coord/`
139	artifacts.
140
141	## Subagent review (keep main session lean)
142
143	When: ≥ 3 task packets exist and you want to verify they stayed
144	within the declared `task_packet_token_budget` (default 6000 tokens
145	≈ 24 KB) without re-reading them all in the main session.
146
147	Why: After writing `.coord/context_<NNN>.md` + `session_primer.md`,
148	the main session has the policy in context but doesn't need to re-read
149	every packet to verify compliance. A subagent can scan all packets
150	and return

People who liked this also viewed

6 picks

Type

agent-context-budget

16stars4forks

View on GitHub

How to install agent-context-budget?

wenyuchiou/agent-collab-skills/agent-context-budget

$npx -y skills add WenyuChiou/agent-collab-skills --skill agent-context-budget

Installs into the current project.

Summary

Use when multi-agent work risks context overflow, memory growth, noisy logs, oversized handoffs, cross-session continuation, or parallel Codex and Gemini execution.

Files · 1

View on GitHub

SKILL.md

1	# agent-context-budget
2
3	Context governor for multi-agent rounds. The core rule is simple:
4	keep `.coord/` as the canonical state, pass agents compact packets,
5	and never paste raw logs or unbounded memory into the main session.
6
7	## When to Use
8
9	Use this before or during:
10
11	- Large Codex + Gemini + Claude runs.
12	- Cross-session resumes where the prior conversation is too large.
13	- Any round with more than two delegate tasks.
14	- Any workflow where `.coord/memory.yml`, `.ai/_log_.txt`, or agent
15	summaries are starting to dominate the prompt.
16
17	Not for small one-agent edits. Use the direct delegate skill instead.
18
19	## Default Policy
20
21	If `.coord/plan.yml` lacks `context_policy`, add this block:
22
23	```yaml
24	context_policy:
25	main_session_token_budget: 3000
26	task_packet_token_budget: 6000
27	result_summary_word_budget: 250
28	memory_digest_token_budget: 1200
29	log_tail_lines_on_error: 50
30	raw_log_policy: path-only
31	agentmemory: optional
32
33	# W3 — $ cost gate per task (v0.2.2+)
34	# Optional. If set, agent-acceptance-gate flags any task whose
35	# codex/gemini delegate exceeded the cap. Estimated from token
36	# usage × provider price (Anthropic / OpenAI / Google rates).
37	# Set per-task in plan.yml tasks[].budget.max_cost_usd to override.
38	default_max_cost_usd: 0.50 # default $ ceiling per task
39	total_round_max_cost_usd: 5.00 # hard stop for the entire round
40	```
41
42	Why both `token` and `cost` budgets? Token gate prevents context
43	bloat (a session quality concern). Cost gate prevents runaway delegation
44	spend (a financial concern). They're orthogonal: a 2k-token delegate
45	call can cost $0.05 (Claude Haiku) or $0.50 (Claude Opus), so token
46	count alone doesn't bound dollars.
47
48	Per-task override example:
49
50	```yaml
51	# in plan.yml
52	tasks:
53	- id: T1
54	agent: codex
55	slug: simple-refactor
56	budget:
57	max_cost_usd: 0.10 # this task is mechanical, cap low
58	- id: T2
59	agent: codex
60	slug: complex-rewrite
61	budget:
62	max_cost_usd: 2.00 # this task needs frontier model, allow higher
63	```
64
65	## Workflow
66
67	1. Read `.coord/plan.yml` and `.coord/memory.yml` if present.
68	2. Write `.coord/context_<NNN>.md` with:
69	- round goal and success criteria
70	- task graph and write ownership
71	- per-agent context packets to include
72	- context intentionally excluded
73	- optional `agentmemory` recall queries, if available
74	3. Write or refresh `.coord/session_primer.md` with:
75	- current decisions
76	- open questions
77	- active round status
78	- recent artifacts by path
79	- no raw logs
80	4. Enforce result packets:
81	- delegate summaries are <= 250 words
82	- changed files are listed by path
83	- tests are listed by command + result
84	- risks are explicit and short
85	5. If a log is needed, include only the path. Read the last 50 lines
86	only when the corresponding `result.json` status is `error`.
87
88	## Optional agentmemory
89
90	`agentmemory` is a recall cache, not the source of truth.
91
92	- Query it only to enrich `.coord/session_primer.md`.
93	- Store only compact memory candidates: accepted decisions, resolved
94	open questions, artifact summaries, and final session outcomes.
95	- If it is unavailable, continue with `.coord/memory.yml` only.
96	- Never use vector recall as acceptance evidence. The acceptance gate
97	must read `.coord/plan.yml`, result files, reconciliation, and tests.
98
99	## Memory Promotion Rules
100
101	Promote only:
102
103	- Decisions with a one-sentence `what` and `why`.
104	- Open questions with a blocker and suggested next owner.
105	- Artifact pointers with one-line summaries.
106	- Agent session outcomes with status and result-summary path.
107
108	Do not promote:
109
110	- Raw logs.
111	- Full diffs.
112	- Source code.
113	- Long analysis.
114	- Secrets or credentials.
115
116	## Output
117
118	End with:
119
120	```text
121	[agent-context-budget]
122	Round: <N>
123	Policy: .coord/plan.yml context_policy
124	Context plan: .coord/context_<NNN>.md
125	Session primer: .coord/session_primer.md
126	Raw logs: path-only; failure tail max 50 lines
127	agentmemory: optional cache, not required
128	```
129
130	## Common Mistakes
131
132	- Pasting full logs into chat. Store the path; read bounded tail only
133	on failure.
134	- Treating memory as a transcript. Promote decisions, questions,
135	artifacts, and outcomes only.
136	- Giving Gemini only `.ai/` paths. Inline critical context because
137	`.ai/` may be gitignored.
138	- Using agentmemory as the gate. Gate with tests and `.coord/`
139	artifacts.
140
141	## Subagent review (keep main session lean)
142
143	When: ≥ 3 task packets exist and you want to verify they stayed
144	within the declared `task_packet_token_budget` (default 6000 tokens
145	≈ 24 KB) without re-reading them all in the main session.
146
147	Why: After writing `.coord/context_<NNN>.md` + `session_primer.md`,
148	the main session has the policy in context but doesn't need to re-read
149	every packet to verify compliance. A subagent can scan all packets
150	and return

People who liked this also viewed

6 picks

Type